Design of a new method for detection of occupancy in the smart home using an FBG sensor

sensors Article Design of a New Method for Detection of Occupancy in the Smart Home Using an FBG Sensor Jan Vanus 1,* , Jan Nedoma 2, Marcel Fajkus 2and Radek Martinek 1 1Department of Cybernetics and Biomedical Engineering, Faculty of Electrical Engineering and Computer Science, VSB–Technical University of Ostrava, 708 33 Ostrava, Czech Republic; [email protected] 2 Department of Telecommunications, Faculty of Electrical Engineering and Computer Science, VSB–Technical University of Ostrava, 708 33 Ostrava, Czech Republic; [email protected] (J.N.); [email protected] (M.F.) *Correspondence: [email protected] Received: 29 November 2019; Accepted: 6 January 2020; Published: 10 January 2020   Abstract: This article introduces a new way of using a fibre Bragg grating (FBG) sensor for detecting the presence and number of occupants in the monitored space in a smart home (SH). CO 2 sensors are used to determine the CO 2 concentration of the monitored rooms in an SH. CO 2 sensors can also be used for occupancy recognition of the monitored spaces in SH. To determine the presence of occupants in the monitored rooms of the SH, the newly devised method of CO 2 prediction, by means of an artificial neural network (ANN) with a scaled conjugate gradient (SCG) algorithm using measurements of typical operational technical quantities (indoor temperature, relative humidity indoor and CO 2 concentration in the SH) is used. The goal of the experiments is to verify the possibility of using the FBG sensor in order to unambiguously detect the number of occupants in the selected room (R104) and, at the same time, to harness the newly proposed method of CO 2 prediction with ANN SCG for recognition of the SH occupancy status and the SH spatial location (rooms R104, R203, and R204) of an occupant. The designed experiments will verify the possibility of using a minimum number of sensors for measuring the non-electric quantities of indoor temperature and indoor relative humidity and the possibility of monitoring the presence of occupants in the SH using CO 2 prediction by means of the ANN SCG method with ANN learning for the data obtained from only one room (R203). The prediction accuracy exceeded 90% in certain experiments. The uniqueness and innovativeness of the described solution lie in the integrated multidisciplinary application of technological procedures (the BACnet technology control SH, FBG sensors) and mathematical methods (ANN prediction with SCG algorithm, the adaptive filtration with an LMS algorithm) employed for the recognition of number persons and occupancy recognition of selected monitored rooms of SH. Keywords: smart home (SH); prediction; artificial neural network (ANN); fiber bragg grating (FBG); occupancy; number of person recognition; scaled conjugate gradient (SCG) 1. Introduction Recognizing the occupancy, number of individuals, location-and-movement recognition and activity recognition of an individual in an indoor space is one the key functionalities of a smart home (SH), as a prerequisite to providing services to support independent living of elderly SH occupants and has a great influence on internal loads and HVAC (heating, ventilation and air conditioning) requirement, thus increasing the energy consumption optimization. Azghandi et al. focused on the particular case of an SH with multiple occupants, they developed a location-and-movement recognition method using many inexpensive passive infrared (PIR) motion sensors and a small number of more Sensors 2020,20, 398; doi:10.3390/s20020398 www.mdpi.com/journal/sensors Sensors 2020,20, 398 2 of 31 costly radio frequency identification (RFID) readers [ 1 ]. Benmansour et al. provided an overview of existing approaches and current practices for activity recognition in multi-occupant SHs [ 2 ]. Braun et al. reported the investigation of two categories of occupancy sensors with the requirements of supporting wireless communication and a focus on the low cost of the systems (capacitive proximity sensors and accelerometers that are placed below the furniture) with a classification accuracy between 79% and 96% [ 3 ]. Chan et al. proposed the methodology and design of a voice-controlled environment, with an emphasis on speech recognition and voice control, based on Amazon Alexa and Raspberry Pi in an SH [ 4 ]. Chen et al. proposed an activity recognition system guided by an unobtrusive sensor (ARGUS) with a facing direction detection accuracy, resulting from manually defined features, that reached 85.3%, 90.6%, and 85.2% [ 5 ]. Khan et al. developed a low-cost heterogeneous radar-based activity monitoring (RAM) system for recognizing fine-grained activities in an SH with detecting accuracy of 92.84% [ 6 ]. Lee et al. investigated the use of cameras and a distributed processing method for the automated control of lights in an SH, which provided occupancy reasoning and human activity analysis [ 7 ]. Mokhtari et al. proposed a new human identification sensor, which can efficiently differentiate multiple residents in a home environment to detect their height as a unique bio-feature with three sensing/communication modules: pyroelectric infrared (PIR) occupancy, ultrasound array, and Bluetooth low-energy (BLE) communication modules [ 8 ]. A new recognition algorithm for household appliances, based on a Bayes classification model, is presented by Yan et al., in which sequential appliance power consumption data from intelligent power sockets is used and for the generalization and extraction of the characteristics of occupant behavior and power consumption of typical household appliances [ 9 ]. Yang et al. proposed a novel indoor tracking technique for SHs with multiple residents by relying only on non-wearable, environmentally deployed sensors such as passive infrared motion sensors [ 10 ]. Feng et al. presented a novel real-time, device-free, and privacy-preserving WiFi-enabled Internet of Things (IoT) platform for SH-occupancy sensing, which can promote a myriad of emerging applications with an accuracy of 96.8% and 90.6% in terms of occupancy detection and recognition, respectively [ 11 ]. Traditionally, in building energy modeling (BEM) programs, occupant behavior (OB) inputs are deterministic and less indicative of real-world scenarios, contributing to discrepancies between simulated and actual energy use in buildings. Yin et al. (2016) presented a new OB modeling tool, with an occupant behavior functional mock-up unit (obFMU) that enables co-simulation with BEM programs implementing a functional mock-up interface (FMI) [ 12 ]. Occupants are involved in a variety of activities in buildings, which drive them to move among rooms, enter or leave a building. Hong et al. (2016) defined SH occupancy using four parameters and showed how they varied with time. The four occupancy parameters were as follows: (1) the number of occupants in a building, (2) occupancy status of space, (3) the number of occupants in a space, and (4) the location of an occupant [13]. In order to detect the occupancy of the SH by indirect methods (without using cameras), common operational and technical sensors are used in this article to measure the indoor temperature, the indoor relative humidity and the CO 2 indoor concentration within the BACnet technology for HVAC control. A fibre Bragg grating (FBG) sensor will be used to detect the number of occupants in the monitored space of room R104 (ground floor). The method devised for the prediction of the CO 2 waveform using artificial neural network (ANN) scaled conjugate gradient (SCG) will be verified during the experiments conducted to detect the occupancy of rooms R104, R203 and R204. The input quantities measured to ANN SCG were obtained from the indoor temperature and relative indoor humidity sensors. One of the objectives of the article is to verify the possibility of minimizing investment costs by using cheaper temperature and relative humidity sensors instead of a more expensive CO 2 sensor to detect the occupancy of monitored SH spaces. The other objectives of this article are the following: 1. Experimental verification of FBG sensor use for the recognition of the number occupants in SH room R104. 2. Experimental verification of the CO 2 concentration measurement in an SH by means of common operational sensors for the occupancy status of the SH space. Sensors 2020,20, 398 3 of 31 3. Experimental verification of the method with ANN SCG that was devised for CO 2 concentration prediction (more one-day measurements in the period from 25 June 2018, to 28 June 2018) to locate an occupant (in rooms R104, R203, and R204) in an SH with the highest possible accuracy. 4. Experimental verification of the possibility of ANN learning for one room only (R203) in order to predict CO2concentrations in other rooms (R104, R204). The experimental measurements of objective parameters of the internal environment and thermal comfort evaluation were conducted in selected SH rooms R204, R203, and R104 in a wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava. (Figure 1) [ 14 ]. Sensors 2020, 20, x FOR PEER REVIEW 3 of 29 3. Experimental verification of the method with ANN SCG that was devised for CO 2 concentration prediction (more one-day measurements in the period from 25 June 2018, to 28 June 2018) to locate an occupant (in rooms R104, R203, and R204) in an SH with the highest possible accuracy. 4. Experimental verification of the possibility of ANN learning for one room only (R203) in order to predict CO 2 concentrations in other rooms (R104, R204). The experimental measurements of objective parameters of the internal environment and thermal comfort evaluation were conducted in selected SH rooms R204, R203, and R104 in a wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava. (Figure 1) [14]. Figure 1. The wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava with selected smart home (SH) rooms R204, R203, and R104 [15]. 2. Materials and Methods 2.1. Fiber Bragg Grating (FBG) Sensor Using for Recognition of Number Occupants in Smart Home (SH) Room R104 Bragg gratings (FBG) are special structures created by an ultraviolet (UV) laser inside the core of a photosensitive optical fibre. This structure consists of a periodic structure of changes in the refractive index, where the layers of the refractive index of the core 𝑛 alternate with the layers of the increased refractive index 𝑛 (1): 𝑛=𝑛  +𝛿𝑛 (1) where 𝛿𝑛 is the refractive index induced by UV radiation [16]. When a broad-spectrum light is introduced into the optical fibre, the Bragg grating reflects a narrow spectral portion and all the other wavelengths pass through the structure without damping (Figure 2). Figure 2. The Bragg grating principle. Figure 1. The wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava with selected smart home (SH) rooms R204, R203, and R104 [15]. 2. Materials and Methods 2.1. Fiber Bragg Grating (FBG) Sensor Using for Recognition of Number Occupants in Smart Home (SH) Room R104 Bragg gratings (FBG) are special structures created by an ultraviolet (UV) laser inside the core of a photosensitive optical fibre. This structure consists of a periodic structure of changes in the refractive index, where the layers of the refractive index of the core n1 alternate with the layers of the increased refractive index n3(1): n3=n1+δn(1) where δnis the refractive index induced by UV radiation [16]. When a broad-spectrum light is introduced into the optical fibre, the Bragg grating reflects a narrow spectral portion and all the other wavelengths pass through the structure without damping (Figure 2). Sensors 2020, 20, x FOR PEER REVIEW 3 of 29 3. Experimental verification of the method with ANN SCG that was devised for CO 2 concentration prediction (more one-day measurements in the period from 25 June 2018, to 28 June 2018) to locate an occupant (in rooms R104, R203, and R204) in an SH with the highest possible accuracy. 4. Experimental verification of the possibility of ANN learning for one room only (R203) in order to predict CO 2 concentrations in other rooms (R104, R204). The experimental measurements of objective parameters of the internal environment and thermal comfort evaluation were conducted in selected SH rooms R204, R203, and R104 in a wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava. (Figure 1) [14]. Figure 1. The wooden building of the passive standard located in the Faculty of Civil Engineering, VSB—TU Ostrava with selected smart home (SH) rooms R204, R203, and R104 [15]. 2. Materials and Methods 2.1. Fiber Bragg Grating (FBG) Sensor Using for Recognition of Number Occupants in Smart Home (SH) Room R104 Bragg gratings (FBG) are special structures created by an ultraviolet (UV) laser inside the core of a photosensitive optical fibre. This structure consists of a periodic structure of changes in the refractive index, where the layers of the refractive index of the core 𝑛 alternate with the layers of the increased refractive index 𝑛 (1): 𝑛=𝑛  +𝛿𝑛 (1) where 𝛿𝑛 is the refractive index induced by UV radiation [16]. When a broad-spectrum light is introduced into the optical fibre, the Bragg grating reflects a narrow spectral portion and all the other wavelengths pass through the structure without damping (Figure 2). Figure 2. The Bragg grating principle. Figure 2. The Bragg grating principle. Sensors 2020,20, 398 4 of 31 The central wavelength of the reflected spectral portion is called the Bragg wavelength λB and is defined by the optical and geometric properties of the structure according to (2): λB=2nef f Λ(2) where nef f is the effective refractive index of the periodic structure and Λ is the distance between the periodic changes in the refractive index. The external effects of the temperature and the deformation influence the optical and geometric properties and thus, the spectral position of the Bragg wavelength. Thanks to this feature, Bragg gratings are used in sensory applications. The dependence of the Bragg wavelength on the deformation and the temperature is expressed by (3): ∆λB λB =kε+ (αΛ+αn)∆T(3) where k is the deformation coefficient, ε is the optical fibre deformation caused by measurement, αΛ is the coefficient of thermal expansion, αn is the thermo-optic coefficient and ∆T is the change in the operating temperature [17]. Bragg gratings in a standard optical fibre with a central wavelength of 1550 nm show a deformation sensitivity of 1.1 pm/ µ strain and a temperature sensitivity of 10.3 pm/ ◦ C. By using a suitable encapsulation, it is possible to implement a sensor of almost any physical quantity. FBG sensors are used in automobile [ 18 ] and railway transport [ 19 ], the construction industry [ 20 ], power engineering, biomedical [21] or perimetric applications [22], etc. Bragg gratings are single-point sensors. By using a wavelength or time multiplex, it is possible to connect tens or hundreds of these sensors in a single optical fibre in order to achieve a quasi-distributed sensory system [23]. 2.1.1. Fiberglass Bragg Sensors One of the most widespread applications of Bragg gratings includes deformation and compression measurements. Depending on the type of application, Bragg gratings can be encapsulated in many ways. The principle of the encapsulation is to protect the fragile glass fibre, to enhance the sensitivity to the desired quantity and to suppress the surrounding interference. Bragg gratings can be encapsulated in polymers [ 24 ] of fibreglass or composite materials [ 25 , 26 ], and special steel jigs [ 27 ], which are mechanically attached to the structure that is to be measured, etc. Based on the advantages of using Bragg gratings—such as their reliable and very accurate measurement—these types of sensors were used for the reference measurement of the occupancy recognition of room R104 in the SH. Because the grating sensors were installed on a wooden staircase, the encapsulation of Bragg gratings in fibreglass strips was used. This method enables the implementation of a very thin sensor that transmits deformations from the step (passage of persons) to the optical fibre itself. Two Bragg gratings in a single-mode optical fibre with primary acrylate protection were used for implementing the sensors. Bragg gratings A and B had the following parameters: The Bragg wavelengths were 1547.510 nm and 1552.369 nm, respectively; the reflection spectrum width was 256 pm and 227 pm, respectively; the reflectivity was 91.2% and 91.3%, respectively. Each Bragg grating was placed between the glass fabrics (2 layers below and 2 layers above the optical fibre). The glass fabric was then coated with a polymer resin. The actual curing caused a Bragg wavelength shift of 23 µ m for Sensor A and of 19 µ m for Sensor B to lower wavelengths (Figure 3). The Bragg grating is located in the middle of the fibreglass strip, marked in red. Sensors 2020,20, 398 5 of 31 Sensors 2020, 20, x FOR PEER REVIEW 4 of 29 The central wavelength of the reflected spectral portion is called the Bragg wavelength 𝜆 and is defined by the optical and geometric properties of the structure according to (2): 𝜆=2𝑛Λ (2) where 𝑛 is the effective refractive index of the periodic structure and Λ is the distance between the periodic changes in the refractive index. The external effects of the temperature and the deformation influence the optical and geometric properties and thus, the spectral position of the Bragg wavelength. Thanks to this feature, Bragg gratings are used in sensory applications. The dependence of the Bragg wavelength on the deformation and the temperature is expressed by (3): Δ𝜆 𝜆=𝑘𝜀+(𝛼+𝛼)Δ𝑇 (3) where 𝑘 is the deformation coefficient, 𝜀 is the optical fibre deformation caused by measurement, 𝛼 is the coefficient of thermal expansion, 𝛼 is the thermo-optic coefficient and Δ𝑇 is the change in the operating temperature [17]. Bragg gratings in a standard optical fibre with a central wavelength of 1550 nm show a deformation sensitivity of 1.1 pm/µstrain and a temperature sensitivity of 10.3 pm/°C. By using a suitable encapsulation, it is possible to implement a sensor of almost any physical quantity. FBG sensors are used in automobile [18] and railway transport [19], the construction industry [20], power engineering, biomedical [21] or perimetric applications [22], etc. Bragg gratings are single-point sensors. By using a wavelength or time multiplex, it is possible to connect tens or hundreds of these sensors in a single optical fibre in order to achieve a quasidistributed sensory system [23]. 2.1.1. Fiberglass Bragg Sensors One of the most widespread applications of Bragg gratings includes deformation and compression measurements. Depending on the type of application, Bragg gratings can be encapsulated in many ways. The principle of the encapsulation is to protect the fragile glass fibre, to enhance the sensitivity to the desired quantity and to suppress the surrounding interference. Bragg gratings can be encapsulated in polymers [24] of fibreglass or composite materials [25,26], and special steel jigs [27], which are mechanically attached to the structure that is to be measured, etc. Based on the advantages of using Bragg gratings—such as their reliable and very accurate measurement— these types of sensors were used for the reference measurement of the occupancy recognition of room R104 in the SH. Because the grating sensors were installed on a wooden staircase, the encapsulation of Bragg gratings in fibreglass strips was used. This method enables the implementation of a very thin sensor that transmits deformations from the step (passage of persons) to the optical fibre itself. Two Bragg gratings in a single-mode optical fibre with primary acrylate protection were used for implementing the sensors. Bragg gratings A and B had the following parameters: The Bragg wavelengths were 1547.510 nm and 1552.369 nm, respectively; the reflection spectrum width was 256 pm and 227 pm, respectively; the reflectivity was 91.2% and 91.3%, respectively. Each Bragg grating was placed between the glass fabrics (2 layers below and 2 layers above the optical fibre). The glass fabric was then coated with a polymer resin. The actual curing caused a Bragg wavelength shift of 23 µm for Sensor A and of 19 µm for Sensor B to lower wavelengths (Figure 3). The Bragg grating is located in the middle of the fibreglass strip, marked in red. (a) (b) Figure 3. Implementation of the sensor by encapsulating the Bragg grating in fibreglass ( a ); the resulting fibre Bragg grating (FBG) fibreglass sensor (b). 2.1.2. Implementation of FBG Sensors FBG sensors were implemented on the staircase leading from the ground floor to the first floor (Figure 4). The sensors were glued with cyanoacrylate adhesive to the bottom of the second step (FBG A) and the third step (FBG B). Sensors 2020, 20, x FOR PEER REVIEW 5 of 29 Figure 3. Implementation of the sensor by encapsulating the Bragg grating in fibreglass (a); the resulting fibre Bragg grating (FBG) fibreglass sensor (b). 2.1.2. Implementation of FBG Sensors FBG sensors were implemented on the staircase leading from the ground floor to the first floor (Figure 4). The sensors were glued with cyanoacrylate adhesive to the bottom of the second step (FBG A) and the third step (FBG B). Figure 4. Placement of FBG sensors on the staircase in the smart home (SH), room R104. 2.2. Use of a CO 2 Sensor Network for Monitoring SH Space Occupancy Common CO 2 sensors can be used to detect the occupancy of individual SH spaces. In room R104, the BT 12.09 sensor (Figure 5), in room R203 the BT 12.10 sensor and in room R204 the BT 12.10 sensor (Figure 6) were used for measuring the CO 2 concentration in the framework of forced Air Condition (AC) control (Figure 7). The BACnet technology is used in the SH to control HVAC (Figure 8). The presence of occupants in the SH can be detected by measuring the CO 2 concentration. Figure 5 shows the ground plan of the ground floor of the SH with the location of the individual sensors for CO 2 measurement. Figure 5. The ground plan of the ground floor of the SH with the location of the sensors for CO 2 measurement. The individual rooms on the ground floor of the SH are marked as follows (Figure 5): • R101—door space, entrance hall, • R102—toilet 1, Figure 4. Placement of FBG sensors on the staircase in the smart home (SH), room R104. 2.2. Use of a CO2Sensor Network for Monitoring SH Space Occupancy Common CO 2 sensors can be used to detect the occupancy of individual SH spaces. In room R104, the BT 12.09 sensor (Figure 5), in room R203 the BT 12.10 sensor and in room R204 the BT 12.10 sensor (Figure 6) were used for measuring the CO 2 concentration in the framework of forced Air Condition (AC) control (Figure 7). The BACnet technology is used in the SH to control HVAC (Figure 8). The presence of occupants in the SH can be detected by measuring the CO 2 concentration. Figure 5 shows the ground plan of the ground floor of the SH with the location of the individual sensors for CO2measurement. Sensors 2020,20, 398 6 of 31 Sensors 2020, 20, x FOR PEER REVIEW 5 of 29 Figure 3. Implementation of the sensor by encapsulating the Bragg grating in fibreglass (a); the resulting fibre Bragg grating (FBG) fibreglass sensor (b). 2.1.2. Implementation of FBG Sensors FBG sensors were implemented on the staircase leading from the ground floor to the first floor (Figure 4). The sensors were glued with cyanoacrylate adhesive to the bottom of the second step (FBG A) and the third step (FBG B). Figure 4. Placement of FBG sensors on the staircase in the smart home (SH), room R104. 2.2. Use of a CO 2 Sensor Network for Monitoring SH Space Occupancy Common CO 2 sensors can be used to detect the occupancy of individual SH spaces. In room R104, the BT 12.09 sensor (Figure 5), in room R203 the BT 12.10 sensor and in room R204 the BT 12.10 sensor (Figure 6) were used for measuring the CO 2 concentration in the framework of forced Air Condition (AC) control (Figure 7). The BACnet technology is used in the SH to control HVAC (Figure 8). The presence of occupants in the SH can be detected by measuring the CO 2 concentration. Figure 5 shows the ground plan of the ground floor of the SH with the location of the individual sensors for CO 2 measurement. Figure 5. The ground plan of the ground floor of the SH with the location of the sensors for CO 2 measurement. The individual rooms on the ground floor of the SH are marked as follows (Figure 5): • R101—door space, entrance hall, • R102—toilet 1, Figure 5. The ground plan of the ground floor of the SH with the location of the sensors for CO2measurement. Sensors 2020, 20, x FOR PEER REVIEW 6 of 29 • R103—toilet 2, • R104—entrance room; FBG sensor is placed on the staircase, • R105—utility room, there are heating sources, • R106—classroom. Figure 6 shows the ground plan of the first floor of the SH with the location of the individual sensors for CO 2 measurement. Figure 6. The ground plan of the first floor of the SH with the location of the sensors for CO 2 measurement. The individual rooms on the SH first floor are marked as follows (Figure 6): • R201—staircase, • R202—control room, • R203—classroom (office), • R204—classroom (office), • R205—toilet and bathroom. The list (legend) of the individual sensors used (Figures 5–7): • BT 12.01—Measurement at the fresh outdoor air inlet into QPA 2062 SH. • BT 12.02—Measurement at the recirculation air inlet from SH spaces into QPM 2162 heat recovery unit. • BT 12.03—Measurement at the recirculation air inlet from SH spaces into QPA 2062 heat recovery unit. • BT 12.04—Measurement in QFA 2060 heat recovery unit. • BT 12.05—Measurement at the recirculation and fresh air outlet from the heat recovery unit into QFM 2160 SH. • BT 12.06—Measurement at the exhaust air inlet into QFM 2160 recuperation unit. • BT 12.07—Measurement at the exhaust air outlet from QPA 2062 recuperation unit. • BT 12.08—Measurement at the recirculation air inlet from SH spaces into QPM 2162 heat recovery unit. • BT 12.09—sensor located in room R104, QPA 2062. • BT 12.10—sensor located in room R203, QPA 2062. • BT 12.11—sensor located in room R204, QPA 2062. The technical specification of the individual sensors used: • QPA 20.62 room sensor for measuring the air quality—CO 2 , relative humidity and temperature—with a measurement accuracy: (50 ppm + 2% of the value measured, long-term Figure 6. The ground plan of the first floor of the SH with the location of the sensors for CO2measurement. Sensors 2020, 20, x FOR PEER REVIEW 7 of 29 drift: 5% of the measuring range/5 years (typically). The CO 2 sensor principle is based on nondispersive infrared absorption (NDIR) measurement. • QPM 21.62 channel sensors for air quality—CO 2 , relative humidity, temperature. Measurement accuracy: (50 ppm + 2% of the value measured), long-term drift: 5% of the measuring range/5 years (typically). The CO 2 sensor is based on non-dispersive infrared absorption (NDIR) measurement. • QFA 20.60 room sensor for temperature and relative humidity. Measurement accuracy ± 3% rH in within the comfort range. Application range −15 … +50 °C/0 … 95% rH in (no condensation). • QFM 21.60 channel sensor for relative humidity and temperature. Measurement accuracy ± 3% rH in within the comfortable range. Application range −15 … +60 °C/0 … 95% rH in (no condensation). Figure 7 shows the ventilation distribution technology with the location of the individual CO 2 sensors. Figure 7. The ventilation distribution technology with the location of the individual sensors for CO 2 measurement in an SH. A block diagram containing a description of the individual components and function blocks within the BACnet technology in the SH for HVAC control is shown in Figure 8. Figure 8. Block diagram of the BACnet technology used in an SH for HVAC control. 2.3. The Design of the New Method for CO 2 Prediction The newly devised method for CO 2 prediction from the temperature indoor and relative humidity indoor values measured by means of ANN SCG (multiple one-day measurements) was used for the location of an occupant (in rooms R104, R203 a R204) in SH with the highest possible accuracy. Block diagram of processing the quantities measured in SH for multiple one-day Figure 7. The ventilation distribution technology with the location of the individual sensors for CO 2 measurement in an SH. Sensors 2020,20, 398 7 of 31 Sensors 2020, 20, x FOR PEER REVIEW 7 of 29 drift: 5% of the measuring range/5 years (typically). The CO 2 sensor principle is based on nondispersive infrared absorption (NDIR) measurement. • QPM 21.62 channel sensors for air quality—CO 2 , relative humidity, temperature. Measurement accuracy: (50 ppm + 2% of the value measured), long-term drift: 5% of the measuring range/5 years (typically). The CO 2 sensor is based on non-dispersive infrared absorption (NDIR) measurement. • QFA 20.60 room sensor for temperature and relative humidity. Measurement accuracy ± 3% rH in within the comfort range. Application range −15 … +50 °C/0 … 95% rH in (no condensation). • QFM 21.60 channel sensor for relative humidity and temperature. Measurement accuracy ± 3% rH in within the comfortable range. Application range −15 … +60 °C/0 … 95% rH in (no condensation). Figure 7 shows the ventilation distribution technology with the location of the individual CO 2 sensors. Figure 7. The ventilation distribution technology with the location of the individual sensors for CO 2 measurement in an SH. A block diagram containing a description of the individual components and function blocks within the BACnet technology in the SH for HVAC control is shown in Figure 8. Figure 8. Block diagram of the BACnet technology used in an SH for HVAC control. 2.3. The Design of the New Method for CO 2 Prediction The newly devised method for CO 2 prediction from the temperature indoor and relative humidity indoor values measured by means of ANN SCG (multiple one-day measurements) was used for the location of an occupant (in rooms R104, R203 a R204) in SH with the highest possible accuracy. Block diagram of processing the quantities measured in SH for multiple one-day Figure 8. Block diagram of the BACnet technology used in an SH for HVAC control. The individual rooms on the ground floor of the SH are marked as follows (Figure 5): •R101—door space, entrance hall, •R102—toilet 1, •R103—toilet 2, •R104—entrance room; FBG sensor is placed on the staircase, •R105—utility room, there are heating sources, •R106—classroom. Figure 6shows the ground plan of the first floor of the SH with the location of the individual sensors for CO2measurement. The individual rooms on the SH first floor are marked as follows (Figure 6): •R201—staircase, •R202—control room, •R203—classroom (office), •R204—classroom (office), •R205—toilet and bathroom. The list (legend) of the individual sensors used (Figures 5–7): •BT 12.01—Measurement at the fresh outdoor air inlet into QPA 2062 SH. • BT 12.02—Measurement at the recirculation air inlet from SH spaces into QPM 2162 heat recovery unit. • BT 12.03—Measurement at the recirculation air inlet from SH spaces into QPA 2062 heat recovery unit. •BT 12.04—Measurement in QFA 2060 heat recovery unit. • BT 12.05—Measurement at the recirculation and fresh air outlet from the heat recovery unit into QFM 2160 SH. •BT 12.06—Measurement at the exhaust air inlet into QFM 2160 recuperation unit. •BT 12.07—Measurement at the exhaust air outlet from QPA 2062 recuperation unit. • BT 12.08—Measurement at the recirculation air inlet from SH spaces into QPM 2162 heat recovery unit. •BT 12.09—sensor located in room R104, QPA 2062. •BT 12.10—sensor located in room R203, QPA 2062. •BT 12.11—sensor located in room R204, QPA 2062. The technical specification of the individual sensors used: Sensors 2020,20, 398 8 of 31 • QPA 20.62 room sensor for measuring the air quality—CO 2 , relative humidity and temperature—with a measurement accuracy: (50 ppm +2% of the value measured, long-term drift: 5% of the measuring range/5 years (typically). The CO 2 sensor principle is based on non-dispersive infrared absorption (NDIR) measurement. • QPM 21.62 channel sensors for air quality—CO 2 , relative humidity, temperature. Measurement accuracy: (50 ppm +2% of the value measured), long-term drift: 5% of the measuring range/5 years (typically). The CO 2 sensor is based on non-dispersive infrared absorption (NDIR) measurement. • QFA 20.60 room sensor for temperature and relative humidity. Measurement accuracy ± 3% rH in within the comfort range. Application range −15 . . . +50 ◦C/0. . . 95% rHin (no condensation). • QFM 21.60 channel sensor for relative humidity and temperature. Measurement accuracy ± 3% rH in within the comfortable range. Application range − 15 . . . +60 ◦ C/0 . . . 95% rH in (no condensation). Figure 7shows the ventilation distribution technology with the location of the individual CO2sensors. A block diagram containing a description of the individual components and function blocks within the BACnet technology in the SH for HVAC control is shown in Figure 8. 2.3. The Design of the New Method for CO2Prediction The newly devised method for CO 2 prediction from the temperature indoor and relative humidity indoor values measured by means of ANN SCG (multiple one-day measurements) was used for the location of an occupant (in rooms R104, R203 a R204) in SH with the highest possible accuracy. Block diagram of processing the quantities measured in SH for multiple one-day measurements in the period from 25 June 2018, to 28 June 2018, using the method devised for CO 2 prediction by means of ANN SCG is shown in Figure 9. Sensors 2020, 20, x FOR PEER REVIEW 8 of 29 measurements in the period from 25 June 2018, to 28 June 2018, using the method devised for CO 2 prediction by means of ANN SCG is shown in Figure 9. Figure 9. Block diagram describing processing of the data measured by means of a scaled conjugate gradient artificial neural network (ANN SCG) within the method devised for CO 2 prediction. The measured values that were used in these experiments are the indoor CO 2 concentration, indoor relative humidity and indoor temperature. The data were pre-processed to improve the efficiency of neural network training. That means that data were normalized so that all the values are between 0 and 1 (Figure 10, Step 1). Figure 10. Block scheme summarizing the experiment steps. In Matlab, the function nftool (neural fitting) was used with neurons varying from 10 to 100 and the three methods previously mentioned. The data samples were divided into 3 sets: training (used to teach the network), validation and testing (provides an independent measure of the network training) (Figure 10, Step 2). After training the networks with data measured on 25 July for room 203, the 90 networks were used to predict data for the rest of the dates from the same room, as well as the two others. Since the learning date was different from the prediction dates, this was called “crossvalidation”. In this step, the function used in Matlab was ‘nntool’ (Figure 10, Step 3). Once the results were ready, the next step was to calculate some parameters that would allow us to quantify the precision of the results, and, therefore, compare the prediction quality between the different ANN SCG (Figure 10, Step 4). The three parameters we relied on for our experiments are as follows. R (correlation coefficient) is a statistical measure that calculates the strength of the relationship between the relative movements of two variables and is calculated with the formula (4). The values range between −1 and 1. A value of exactly 1.0 means there is a perfect positive relationship between the two variables. For a positive increase in one variable, there is also a positive increase in the second variable. A value of −1.0 means there is a perfect negative relationship between the two variables. This shows that the variables move in opposite directions—for a positive increase in one variable, there is a decrease in the second variable. If the correlation is 0, there is no relationship between the two variables [28]. 𝑅= ∑(𝑥−𝑥)(𝑦−𝑦) ∑(𝑥−𝑥)∑(𝑦−𝑦) (4) Figure 9. Block diagram describing processing of the data measured by means of a scaled conjugate gradient artificial neural network (ANN SCG) within the method devised for CO2prediction. The measured values that were used in these experiments are the indoor CO 2 concentration, indoor relative humidity and indoor temperature. The data were pre-processed to improve the efficiency of neural network training. That means that data were normalized so that all the values are between 0 and 1 (Figure 10, Step 1). Sensors 2020,20, 398 9 of 31 Sensors 2020, 20, x FOR PEER REVIEW 8 of 29 measurements in the period from 25 June 2018, to 28 June 2018, using the method devised for CO 2 prediction by means of ANN SCG is shown in Figure 9. Figure 9. Block diagram describing processing of the data measured by means of a scaled conjugate gradient artificial neural network (ANN SCG) within the method devised for CO 2 prediction. The measured values that were used in these experiments are the indoor CO 2 concentration, indoor relative humidity and indoor temperature. The data were pre-processed to improve the efficiency of neural network training. That means that data were normalized so that all the values are between 0 and 1 (Figure 10, Step 1). Figure 10. Block scheme summarizing the experiment steps. In Matlab, the function nftool (neural fitting) was used with neurons varying from 10 to 100 and the three methods previously mentioned. The data samples were divided into 3 sets: training (used to teach the network), validation and testing (provides an independent measure of the network training) (Figure 10, Step 2). After training the networks with data measured on 25 July for room 203, the 90 networks were used to predict data for the rest of the dates from the same room, as well as the two others. Since the learning date was different from the prediction dates, this was called “crossvalidation”. In this step, the function used in Matlab was ‘nntool’ (Figure 10, Step 3). Once the results were ready, the next step was to calculate some parameters that would allow us to quantify the precision of the results, and, therefore, compare the prediction quality between the different ANN SCG (Figure 10, Step 4). The three parameters we relied on for our experiments are as follows. R (correlation coefficient) is a statistical measure that calculates the strength of the relationship between the relative movements of two variables and is calculated with the formula (4). The values range between −1 and 1. A value of exactly 1.0 means there is a perfect positive relationship between the two variables. For a positive increase in one variable, there is also a positive increase in the second variable. A value of −1.0 means there is a perfect negative relationship between the two variables. This shows that the variables move in opposite directions—for a positive increase in one variable, there is a decrease in the second variable. If the correlation is 0, there is no relationship between the two variables [28]. 𝑅= ∑(𝑥−𝑥)(𝑦−𝑦) ∑(𝑥−𝑥)∑(𝑦−𝑦) (4) Figure 10. Block scheme summarizing the experiment steps. In Matlab, the function nftool (neural fitting) was used with neurons varying from 10 to 100 and the three methods previously mentioned. The data samples were divided into 3 sets: training (used to teach the network), validation and testing (provides an independent measure of the network training) (Figure 10, Step 2). After training the networks with data measured on 25 July for room 203, the 90 networks were used to predict data for the rest of the dates from the same room, as well as the two others. Since the learning date was different from the prediction dates, this was called “cross-validation”. In this step, the function used in Matlab was ‘nntool’ (Figure 10, Step 3). Once the results were ready, the next step was to calculate some parameters that would allow us to quantify the precision of the results, and, therefore, compare the prediction quality between the different ANN SCG (Figure 10, Step 4). The three parameters we relied on for our experiments are as follows. R (correlation coefficient) is a statistical measure that calculates the strength of the relationship between the relative movements of two variables and is calculated with the formula (4). The values range between − 1 and 1. A value of exactly 1.0 means there is a perfect positive relationship between the two variables. For a positive increase in one variable, there is also a positive increase in the second variable. A value of − 1.0 means there is a perfect negative relationship between the two variables. This shows that the variables move in opposite directions—for a positive increase in one variable, there is a decrease in the second variable. If the correlation is 0, there is no relationship between the two variables [28]. R=P(x−x)(y−y) qP(x−x)2P(y−y)2 (4) The MSE (mean squared error) parameter describes how close a regression line is to a set of points and is calculated with formula (5). It does this by taking the distances from the points to the regression line (these distances are the “errors”) and squaring them. The squaring is necessary to remove any negative signs. It also gives more weight to larger differences. It is called the mean squared error as you are finding the average of a set of errors [29]: MSE =1 n n X i=1yi−y∗ i(5) MAPE (average absolute percentage error) is a statistical measurement parameter of how accurate a forecast system is. It measures this accuracy as a percentage, and it can be calculated as the average absolute percent error for each time period minus actual values divided by actual values which are given by (6) [30]: MAPE =1 n n X i=1 yi−y∗ i y∗ i (6) yi: reference value, y∗ i: predicted value, n: total number of values, x,y: mean of x, y. After calculating the MAPE, MSE, and R correlation parameters, we plot two figures. The first one has a reference and predicted CO 2 over time (Figures 18, 20 and 22) and the second one is a Sensors 2020,20, 398 16 of 31 indoor temperature (T in ) and indoor relative humidity (rH in ) sensors. The experiments performed within the newly devised method are presented in the following text. 3.3. Experimental Verification of the Method with the Devised Artificial Neural Network (ANN) Scaled Conjugate Gradient (SCG) The conditions of experiments 3.3a and 3.3b were as follows. The experiments were performed for the waveforms of temperature (T in ), relative humidity (rH in ) and CO 2 concentration measured on 26 June 2018, 27 June 2018, and 28 June 2018, (step 1) for rooms R104, R203 and R204 (Figure 17). Sensors 2020, 20, x FOR PEER REVIEW 15 of 29 Figure 16. Waveforms of the CO 2 concentration values measured in SH room R204 (BT 12.11). 3.3. Experimental Verification of the Method with the Devised Artificial Neural Network (ANN) Scaled Conjugate Gradient (SCG) The conditions of experiments 3.3a and 3.3b were as follows. The experiments were performed for the waveforms of temperature (T in ), relative humidity (rH in ) and CO 2 concentration measured on 26 June 2018, 27 June 2018, and 28 June 2018, (step 1) for rooms R104, R203 and R204 (Figure 17). Figure 17. The waveform of the experiments performed on 26 June 2018, 27 June 2018, and 28 June 2018 for rooms R104, R203 and R204. Furthermore, the measured data were pre-processed (step 2) (Figure 17). The design of the prediction system structure (step 3), the implementation (step 4) and ANN SCG training (step 5) for CO 2 prediction using two input values of temperature (T in ) and relative humidity (rH in ) were carried Figure 17. The waveform of the experiments performed on 26 June 2018, 27 June 2018, and 28 June 2018 for rooms R104, R203 and R204. Furthermore, the measured data were pre-processed (step 2) (Figure 17). The design of the prediction system structure (step 3), the implementation (step 4) and ANN SCG training (step 5) for CO 2 prediction using two input values of temperature (T in ) and relative humidity (rH in ) were carried out. The ANN SCG structure designed (Figure 18) for experiment 3.3a was trained (step 5) for the temperature (T in ), relative humidity (rH in ) and CO 2 concentration values measured in room R203. The actual CO 2 prediction (step 6) was implemented for the data measured in rooms R104, R203 and R204. To increase the accuracy of the newly devised method, an additional quantity from the FBG sensor containing the number of occupants in the monitored space of R104 was added to the original two input quantities. The ANN SCG structure designed for experiment 3.3b for CO 2 prediction using three input quantities—temperature (T in ), relative humidity (rH in ) and the number of occupants (FBG sensor). The procedure for performing the experiments was the same as that described in Figure 17. Sensors 2020,20, 398 17 of 31 Sensors 2020, 20, x FOR PEER REVIEW 16 of 29 out. The ANN SCG structure designed (Figure 18) for experiment 3.3a was trained (step 5) for the temperature (T in ), relative humidity (rH in ) and CO 2 concentration values measured in room R203. The actual CO 2 prediction (step 6) was implemented for the data measured in rooms R104, R203 and R204. To increase the accuracy of the newly devised method, an additional quantity from the FBG sensor containing the number of occupants in the monitored space of R104 was added to the original two input quantities. The ANN SCG structure designed for experiment 3.3b for CO 2 prediction using three input quantities—temperature (T in ), relative humidity (rH in ) and the number of occupants (FBG sensor). The procedure for performing the experiments was the same as that described in Figure 17. Experiment 3.3a: Input values, T in and rH in , to ANN SCG for prediction of CO 2 in rooms R203, R204, and R104 (Figure 18). Prediction of CO 2 in rooms R203, R204, R104 for the dates 26, 27 and 28 June 2018 with inputs T in and rH in , using learned ANN SCG (Figure 18) from 25 June 2018. Tables 2–4 show R, MSE and MAPE parameter values, followed by plots of reference and predicted CO 2 , as well as Bland– Altmann plots in rooms 203 (Figures 19 and 20), 204 (Figures 21 and 22) and 104 (Figures 23 and 24). Figure 18. The architecture of the designed ANN SCG on test data measured in R203 from 25 June 2018 for two inputs T in and rH in without and FBG sensor for person presence measuring (PPM). Table 2. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with crossvalidation for 26 June 2018 in R203 (with T in and rH in ), 27 June 2018 in R203 (with T in and rH in ), 28 June 2018 in R203 (with T in and rH in ). 26 June 2018 in R203 27 June 2018 in R203 28 June 2018 in R203 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0062 0.8867 0.2858 0.006 0.5094 0.198 0.0206 0.6398 0.3275 20 0.0063 0.8846 0.289 0.006 0.5098 0.1992 0.0204 0.6449 0.3238 30 0.0058 0.8943 0.2872 0.0058 0.5355 0.1794 0.0198 0.6586 0.3265 40 0.0071 0.8728 0.3237 0.0056 0.5555 0.2036 0.0193 0.6687 0.3186 50 0.0047 0.9151 0.2171 0.0056 0.5591 0.1633 0.0138 0.7784 0.212 60 0.006 0.8918 0.3298 0.0061 0.5084 0.1906 0.0103 0.8401 0.1818 70 0.0059 0.8957 0.3025 0.0054 0.5824 0.1861 0.0127 0.7983 0.2091 80 0.0048 0.9145 0.2475 0.0059 0.5261 0.1845 0.0127 0.7988 0.2558 90 0.0069 0.877 0.1623 0.0065 0.4645 0.2097 0.0106 0.8346 0.2221 100 0.0047 0.916 0.2559 0.0047 0.6489 0.1614 0.0111 0.8266 0.1884 Figure 18. The architecture of the designed ANN SCG on test data measured in R203 from 25 June 2018 for two inputs Tin and rHin without and FBG sensor for person presence measuring (PPM). Experiment 3.3a: Input values, T in and rH in , to ANN SCG for prediction of CO 2 in rooms R203, R204, and R104 (Figure 18). Prediction of CO 2 in rooms R203, R204, R104 for the dates 26, 27 and 28 June 2018 with inputs T in and rH in , using learned ANN SCG (Figure 18) from 25 June 2018. Tables 2–4show R, MSE and MAPE parameter values, followed by plots of reference and predicted CO 2 , as well as Bland–Altmann plots in rooms 203 (Figures 19 and 20), 204 (Figures 21 and 22) and 104 (Figures 23 and 24). Table 2. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with cross-validation for 26 June 2018 in R203 (with T in and rH in ), 27 June 2018 in R203 (with T in and rH in ), 28 June 2018 in R203 (with Tin and rHin). 26 June 2018 in R203 27 June 2018 in R203 28 June 2018 in R203 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0062 0.8867 0.2858 0.006 0.5094 0.198 0.0206 0.6398 0.3275 20 0.0063 0.8846 0.289 0.006 0.5098 0.1992 0.0204 0.6449 0.3238 30 0.0058 0.8943 0.2872 0.0058 0.5355 0.1794 0.0198 0.6586 0.3265 40 0.0071 0.8728 0.3237 0.0056 0.5555 0.2036 0.0193 0.6687 0.3186 50 0.0047 0.9151 0.2171 0.0056 0.5591 0.1633 0.0138 0.7784 0.212 60 0.006 0.8918 0.3298 0.0061 0.5084 0.1906 0.0103 0.8401 0.1818 70 0.0059 0.8957 0.3025 0.0054 0.5824 0.1861 0.0127 0.7983 0.2091 80 0.0048 0.9145 0.2475 0.0059 0.5261 0.1845 0.0127 0.7988 0.2558 90 0.0069 0.877 0.1623 0.0065 0.4645 0.2097 0.0106 0.8346 0.2221 100 0.0047 0.916 0.2559 0.0047 0.6489 0.1614 0.0111 0.8266 0.1884 Sensors 2020,20, 398 18 of 31 Table 3. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with cross-validation for 26 June 2018 in R204 (with T in and rH in ), 27 June 2018 in R204 (with T in and rH in ), and 28 June 2018 in R204 (with Tin and rHin). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0107 0.6276 0.2117 0.0112 0.2603 0.209 0.0187 0.4399 0.1699 20 0.0117 0.5816 0.2634 0.0057 0.7272 0.156 0.0087 0.8007 0.1154 30 0.0116 0.5914 0.2917 0.0049 0.7726 0.1232 0.0186 0.4429 0.1678 40 0.0104 0.6402 0.2231 0.0058 0.7213 0.1535 0.0074 0.825 0.1015 50 0.0103 0.6482 0.2425 0.0061 0.7017 0.1387 0.0186 0.4407 0.1692 60 0.0122 0.5606 0.2589 0.0054 0.7425 0.1463 0.0077 0.817 0.11 70 0.0111 0.6125 0.2503 0.0073 0.6261 0.1668 0.0077 0.8159 0.1065 80 0.0106 0.6354 0.2188 0.0052 0.7551 0.1133 0.0085 0.7971 0.1174 90 0.0111 0.6228 0.1671 0.0052 0.7555 0.1458 0.0077 0.8178 0.1078 100 0.0109 0.6232 0.2742 0.0048 0.7755 0.1404 0.0104 0.7416 0.1202 Table 4. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with cross-validation for 26 June 2018 in R104 (with T in and rH in ), 27 June 2018 in R104 (with T in and rH in ), and 28 June 2018 in R104 (with Tin and rHin). 26 June 2018 in R104 27 June 2018 in R104 28 June 2018 in R104 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0097 0.6714 0.3365 0.0121 0.2188 0.2147 0.0187 0.4399 0.1694 20 0.0102 0.6516 0.2259 0.0052 0.7541 0.1466 0.0186 0.4418 0.1713 30 0.0105 0.6398 0.2233 0.0061 0.7052 0.1395 0.0187 0.4396 0.1699 40 0.0099 0.6643 0.2074 0.0055 0.7389 0.144 0.0128 0.6684 0.1197 50 0.0098 0.6685 0.2099 0.0059 0.7136 0.1319 0.0186 0.4411 0.1692 60 0.0098 0.6697 0.2044 0.0055 0.7387 0.1544 0.0025 0.9135 0.0905 70 0.0099 0.6626 0.1937 0.0056 0.7345 0.1464 0.0082 0.8031 0.1089 80 0.0101 0.6542 0.2063 0.0051 0.7567 0.1134 0.0096 0.7676 0.1121 90 0.0095 0.6822 0.1859 0.0051 0.7613 0.1429 0.0085 0.7959 0.1216 100 0.01 0.6587 0.2005 0.0043 0.8027 0.1202 0.0077 0.8175 0.1041 Sensors 2020, 20, x FOR PEER REVIEW 17 of 29 Figure 19. Comparison of the reference CO 2 concentration waveform and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 20. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Table 3. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with crossvalidation for 26 June 2018 in R204 (with T in and rH in ), 27 June 2018 in R204 (with T in and rH in ), and 28 June 2018 in R204 (with T in and rH in ). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0107 0.6276 0.2117 0.0112 0.2603 0.209 0.0187 0.4399 0.1699 20 0.0117 0.5816 0.2634 0.0057 0.7272 0.156 0.0087 0.8007 0.1154 Figure 19. Comparison of the reference CO 2 concentration waveform and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020,20, 398 19 of 31 Sensors 2020, 20, x FOR PEER REVIEW 17 of 29 Figure 19. Comparison of the reference CO 2 concentration waveform and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 20. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Table 3. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with crossvalidation for 26 June 2018 in R204 (with T in and rH in ), 27 June 2018 in R204 (with T in and rH in ), and 28 June 2018 in R204 (with T in and rH in ). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0107 0.6276 0.2117 0.0112 0.2603 0.209 0.0187 0.4399 0.1699 20 0.0117 0.5816 0.2634 0.0057 0.7272 0.156 0.0087 0.8007 0.1154 Figure 20. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 100 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020, 20, x FOR PEER REVIEW 18 of 29 30 0.0116 0.5914 0.2917 0.0049 0.7726 0.1232 0.0186 0.4429 0.1678 40 0.0104 0.6402 0.2231 0.0058 0.7213 0.1535 0.0074 0.825 0.1015 50 0.0103 0.6482 0.2425 0.0061 0.7017 0.1387 0.0186 0.4407 0.1692 60 0.0122 0.5606 0.2589 0.0054 0.7425 0.1463 0.0077 0.817 0.11 70 0.0111 0.6125 0.2503 0.0073 0.6261 0.1668 0.0077 0.8159 0.1065 80 0.0106 0.6354 0.2188 0.0052 0.7551 0.1133 0.0085 0.7971 0.1174 90 0.0111 0.6228 0.1671 0.0052 0.7555 0.1458 0.0077 0.8178 0.1078 100 0.0109 0.6232 0.2742 0.0048 0.7755 0.1404 0.0104 0.7416 0.1202 Figure 21. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 22. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 21. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020,20, 398 20 of 31 Sensors 2020, 20, x FOR PEER REVIEW 18 of 29 30 0.0116 0.5914 0.2917 0.0049 0.7726 0.1232 0.0186 0.4429 0.1678 40 0.0104 0.6402 0.2231 0.0058 0.7213 0.1535 0.0074 0.825 0.1015 50 0.0103 0.6482 0.2425 0.0061 0.7017 0.1387 0.0186 0.4407 0.1692 60 0.0122 0.5606 0.2589 0.0054 0.7425 0.1463 0.0077 0.817 0.11 70 0.0111 0.6125 0.2503 0.0073 0.6261 0.1668 0.0077 0.8159 0.1065 80 0.0106 0.6354 0.2188 0.0052 0.7551 0.1133 0.0085 0.7971 0.1174 90 0.0111 0.6228 0.1671 0.0052 0.7555 0.1458 0.0077 0.8178 0.1078 100 0.0109 0.6232 0.2742 0.0048 0.7755 0.1404 0.0104 0.7416 0.1202 Figure 21. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 22. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 22. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 40 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020, 20, x FOR PEER REVIEW 19 of 29 Figure 23. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 24. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Table 4. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with crossvalidation for 26 June 2018 in R104 (with T in and rH in ), 27 June 2018 in R104 (with T in and rH in ), and 28 June 2018 in R104 (with T in and rH in ). 26 June 2018 in R104 27 June 2018 in R104 28 June 2018 in R104 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0097 0.6714 0.3365 0.0121 0.2188 0.2147 0.0187 0.4399 0.1694 20 0.0102 0.6516 0.2259 0.0052 0.7541 0.1466 0.0186 0.4418 0.1713 Figure 23. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020,20, 398 21 of 31 Sensors 2020, 20, x FOR PEER REVIEW 19 of 29 Figure 23. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 24. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Table 4. Learned ANN SCG from 25 June 2018 in R203 (with T in and rH in ) and prediction with crossvalidation for 26 June 2018 in R104 (with T in and rH in ), 27 June 2018 in R104 (with T in and rH in ), and 28 June 2018 in R104 (with T in and rH in ). 26 June 2018 in R104 27 June 2018 in R104 28 June 2018 in R104 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0097 0.6714 0.3365 0.0121 0.2188 0.2147 0.0187 0.4399 0.1694 20 0.0102 0.6516 0.2259 0.0052 0.7541 0.1466 0.0186 0.4418 0.1713 Figure 24. Bland–Altman plot for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Experiment 3b: Our next experiment was performed with input values of T in and rH in and presence values from the FBG sensor input into ANN SCG for the prediction of CO 2 in rooms R203, R204, R104 (Figure 25). Prediction of CO 2 in rooms R203, R204, R104 for the dates 26, 27 and 28 June 2018 with input values from sensors T in , rH in and FBG using learned ANN SCG (Figure 25) from 25 June 2018. Tables 5–7 show R, MSE and MAPE parameter values, followed by plots of reference and predicted CO 2 as well as Bland–Altmann plots in rooms 203 (Figures 26 and 27), 204 (Figures 28 and 29) and 104 (Figures 30 and 31). Sensors 2020, 20, x FOR PEER REVIEW 20 of 29 30 0.0105 0.6398 0.2233 0.0061 0.7052 0.1395 0.0187 0.4396 0.1699 40 0.0099 0.6643 0.2074 0.0055 0.7389 0.144 0.0128 0.6684 0.1197 50 0.0098 0.6685 0.2099 0.0059 0.7136 0.1319 0.0186 0.4411 0.1692 60 0.0098 0.6697 0.2044 0.0055 0.7387 0.1544 0.0025 0.9135 0.0905 70 0.0099 0.6626 0.1937 0.0056 0.7345 0.1464 0.0082 0.8031 0.1089 80 0.0101 0.6542 0.2063 0.0051 0.7567 0.1134 0.0096 0.7676 0.1121 90 0.0095 0.6822 0.1859 0.0051 0.7613 0.1429 0.0085 0.7959 0.1216 100 0.01 0.6587 0.2005 0.0043 0.8027 0.1202 0.0077 0.8175 0.1041 Experiment 3b: Our next experiment was performed with input values of T in and rH in and presence values from the FBG sensor input into ANN SCG for the prediction of CO 2 in rooms R203, R204, R104 (Figure 25). Prediction of CO 2 in rooms R203, R204, R104 for the dates 26, 27 and 28 June 2018 with input values from sensors T in , rH in and FBG using learned ANN SCG (Figure 25) from 25 June 2018. Tables 5–7 show R, MSE and MAPE parameter values, followed by plots of reference and predicted CO 2 as well as Bland–Altmann plots in rooms 203 (Figures 26 and 27), 204 (Figures 28 and 29) and 104 (Figures 30 and 31). Figure 25. The architecture of designed ANN SCG on test data measured in R203 from 25 June 2018 for three inputs T in , rH in and those of an FBG sensor for PPM. Table 5. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R203 27 June 2018 in R203 28 June 2018 in R203 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0048 0.9139 0.2285 0.006 0.5094 0.198 0.0201 0.6521 0.2644 20 0.0055 0.9013 0.2814 0.006 0.5098 0.1992 0.0176 0.7052 0.2928 30 0.0057 0.8976 0.301 0.0058 0.5355 0.1794 0.0144 0.7684 0.2296 40 0.0049 0.9117 0.2189 0.0056 0.5555 0.2036 0.0206 0.64 0.3237 50 0.0064 0.8835 0.2802 0.0056 0.5591 0.1633 0.0141 0.7729 0.2079 60 0.004 0.9281 0.2255 0.0061 0.5084 0.1906 0.0132 0.7894 0.2 70 0.0048 0.9135 0.2377 0.0054 0.5824 0.1861 0.0142 0.7704 0.2353 80 0.0052 0.906 0.2541 0.0059 0.5261 0.1845 0.0124 0.8038 0.2117 90 0.0074 0.8684 0.2987 0.0065 0.4645 0.2097 0.0203 0.6467 0.3084 100 0.0044 0.9218 0.1693 0.0047 0.6489 0.1614 0.0203 0.6465 0.3002 Figure 25. The architecture of designed ANN SCG on test data measured in R203 from 25 June 2018 for three inputs Tin, rHin and those of an FBG sensor for PPM. Sensors 2020,20, 398 22 of 31 Table 5. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R203 27 June 2018 in R203 28 June 2018 in R203 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0048 0.9139 0.2285 0.006 0.5094 0.198 0.0201 0.6521 0.2644 20 0.0055 0.9013 0.2814 0.006 0.5098 0.1992 0.0176 0.7052 0.2928 30 0.0057 0.8976 0.301 0.0058 0.5355 0.1794 0.0144 0.7684 0.2296 40 0.0049 0.9117 0.2189 0.0056 0.5555 0.2036 0.0206 0.64 0.3237 50 0.0064 0.8835 0.2802 0.0056 0.5591 0.1633 0.0141 0.7729 0.2079 60 0.004 0.9281 0.2255 0.0061 0.5084 0.1906 0.0132 0.7894 0.2 70 0.0048 0.9135 0.2377 0.0054 0.5824 0.1861 0.0142 0.7704 0.2353 80 0.0052 0.906 0.2541 0.0059 0.5261 0.1845 0.0124 0.8038 0.2117 90 0.0074 0.8684 0.2987 0.0065 0.4645 0.2097 0.0203 0.6467 0.3084 100 0.0044 0.9218 0.1693 0.0047 0.6489 0.1614 0.0203 0.6465 0.3002 Table 6. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0111 0.6096 0.0312 0.0111 0.2745 0.2039 0.0187 0.4407 0.172 20 0.0115 0.5939 0.2684 0.0111 0.302 0.2044 0.0053 0.8785 0.1011 30 0.0106 0.6343 0.2093 0.0063 0.6936 0.1521 0.0186 0.4421 0.1699 40 0.0113 0.6022 0.2116 0.0061 0.702 0.1549 0.0186 0.4422 0.1709 50 0.0107 0.6311 0.2047 0.0058 0.7221 0.1323 0.0082 0.8024 0.1139 60 0.01 0.6595 0.1621 0.0037 0.8352 0.1111 0.0078 0.8139 0.101 70 0.0098 0.6679 0.2093 0.0062 0.6994 0.1459 0.0084 0.7997 0.109 80 0.013 0.5346 0.2705 0.0045 0.7912 0.1289 0.0186 0.4409 0.1695 90 0.0109 0.6201 0.2235 0.0052 0.756 0.1452 0.0048 0.8912 0.0891 100 0.0095 0.6812 0.1807 0.0038 0.8289 0.1018 0.0068 0.8397 0.0948 Sensors 2020,20, 398 23 of 31 Table 7. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R104 27 June 2018 in R104 28 June 2018 in R104 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0104 0.6432 0.2095 0.0112 0.2735 0.2028 0.0231 0.3505 0.2229 20 0.0117 0.5828 0.0638 0.0109 0.3042 0.2001 0.0195 0.4006 0.2088 30 0.0115 0.5987 0.2458 0.006 0.7079 0.151 0.0187 0.4416 0.1723 40 0.0103 0.6482 0.2155 0.0065 0.6818 0.1532 0.0187 0.4404 0.173 50 0.01 0.6612 0.2053 0.0051 0.7592 0.1147 0.0082 0.8034 0.1133 60 0.01 0.6613 0.2 0.0033 0.8534 0.0968 0.0074 0.8245 0.0953 70 0.0114 0.6135 0.2786 0.0066 0.6759 0.1525 0.0083 0.8015 0.108 80 0.0105 0.6379 0.2151 0.0042 0.8063 0.0555 0.0186 0.443 0.1678 90 0.0095 0.6799 0.1926 0.0048 0.7748 0.1293 0.0082 0.8052 0.1112 100 0.0089 0.7067 0.1906 0.0043 0.8026 0.103 0.0076 0.8194 0.1059 Sensors 2020, 20, x FOR PEER REVIEW 21 of 29 Figure 26. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 27. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Table 6. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0111 0.6096 0.0312 0.0111 0.2745 0.2039 0.0187 0.4407 0.172 20 0.0115 0.5939 0.2684 0.0111 0.302 0.2044 0.0053 0.8785 0.1011 Figure 26. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020,20, 398 24 of 31 Sensors 2020, 20, x FOR PEER REVIEW 21 of 29 Figure 26. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 27. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Table 6. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R204 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R204 27 June 2018 in R204 28 June 2018 in R204 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0111 0.6096 0.0312 0.0111 0.2745 0.2039 0.0187 0.4407 0.172 20 0.0115 0.5939 0.2684 0.0111 0.302 0.2044 0.0053 0.8785 0.1011 Figure 27. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 26 June 2018 in R203 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020, 20, x FOR PEER REVIEW 22 of 29 30 0.0106 0.6343 0.2093 0.0063 0.6936 0.1521 0.0186 0.4421 0.1699 40 0.0113 0.6022 0.2116 0.0061 0.702 0.1549 0.0186 0.4422 0.1709 50 0.0107 0.6311 0.2047 0.0058 0.7221 0.1323 0.0082 0.8024 0.1139 60 0.01 0.6595 0.1621 0.0037 0.8352 0.1111 0.0078 0.8139 0.101 70 0.0098 0.6679 0.2093 0.0062 0.6994 0.1459 0.0084 0.7997 0.109 80 0.013 0.5346 0.2705 0.0045 0.7912 0.1289 0.0186 0.4409 0.1695 90 0.0109 0.6201 0.2235 0.0052 0.756 0.1452 0.0048 0.8912 0.0891 100 0.0095 0.6812 0.1807 0.0038 0.8289 0.1018 0.0068 0.8397 0.0948 Figure 28. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 29. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 28. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020,20, 398 25 of 31 Sensors 2020, 20, x FOR PEER REVIEW 22 of 29 30 0.0106 0.6343 0.2093 0.0063 0.6936 0.1521 0.0186 0.4421 0.1699 40 0.0113 0.6022 0.2116 0.0061 0.702 0.1549 0.0186 0.4422 0.1709 50 0.0107 0.6311 0.2047 0.0058 0.7221 0.1323 0.0082 0.8024 0.1139 60 0.01 0.6595 0.1621 0.0037 0.8352 0.1111 0.0078 0.8139 0.101 70 0.0098 0.6679 0.2093 0.0062 0.6994 0.1459 0.0084 0.7997 0.109 80 0.013 0.5346 0.2705 0.0045 0.7912 0.1289 0.0186 0.4409 0.1695 90 0.0109 0.6201 0.2235 0.0052 0.756 0.1452 0.0048 0.8912 0.0891 100 0.0095 0.6812 0.1807 0.0038 0.8289 0.1018 0.0068 0.8397 0.0948 Figure 28. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 29. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 29. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 28 June 2018 in R204 with an ANN with 90 neurons and SCG method trained with data from 25 June 2018 in R203. Sensors 2020, 20, x FOR PEER REVIEW 23 of 29 Figure 30. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 27 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Figure 31. Bland–Altman for the reference and predicted CO 2 waveforms (SRLMS AF) from 27 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203. Table 7. Learned ANN SCG from 25 June 2018 in R203 (with T in, rH in and an FBG sensor for PPM) and prediction with cross-validation for 26 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM), 27 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM), 28 June 2018 in R104 (with T in, rH in and an FBG sensor for PPM). 26 June 2018 in R104 27 June 2018 in R104 28 June 2018 in R104 Number of Neurons ANN SCG MSE R MAPE MSE R MAPE MSE R MAPE (-) (-) (-) (-) (-) (-) (-) (-) (-) 10 0.0104 0.6432 0.2095 0.0112 0.2735 0.2028 0.0231 0.3505 0.2229 20 0.0117 0.5828 0.0638 0.0109 0.3042 0.2001 0.0195 0.4006 0.2088 Figure 30. Comparison of the reference and predicted CO 2 waveforms (SRLMS AF) from 27 June 2018 in R104 with an ANN with 60 neurons and SCG method trained with data from 25 June 2018 in R203.