Predict Household or Grid Electricity Consumption Using Time-Series

International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 300 Predict Household or Grid Electricity Consumption Using Time-Series Prof. Divya H. N.¹, Sneha², Sourabha B. R.³, Shloka S. Kunja⁴, Sai Nandan Sawant⁵ ¹²³⁴⁵ Department of Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bengaluru, Karnataka, India Introduction Household electricity consumption has become increasingly unpredictable due to varying lifestyle patterns, increased appliance usage, and greater dependency on electronic devices. These fluctuations create challenges in balancing Internaonal Journal of Emerging Trends in Engineering and Development Available online on hp://www.rspublica"on.com/ijeted/ijeted_index.htm ISSN 2249-6149 ARTICLE INFO ABSTRACT ©2025 RS Publication Paper ID: IJETED693EC8A3E5C54 Published: 2025-12-15 DOI: https://dx.doi.org/ 10.5281/zenodo.1794 2797 Page No: 300-308 Energy consumption forecasting plays a critical role in optimizing fuel supply planning, resource management, and sustainable energy operations within large institutions such as household. So we focus on medium-term load forecasting for a household, aiming to predict its future energy requirements accurately and efficiently. The primary objective is to support effective consumer electricity planning and energy management strategies. To achieve this, we employ a Temporal Convolutional Network (TCN) model—an advanced deep learning architecture that utilizes one-dimensional convolutional kernels arranged in multiple layers, followed by pooling and fully connected neural network blocks. By leveraging convolutional operations instead of recurrent mechanisms, TCN effectively handles long-range dependencies while maintaining stable gradients, resulting in faster training and better scalability compared to traditional models. For performance evaluation, the TCN model is compared against two widely used time-series forecasting approaches: Long Short-Term Memory (LSTM) networks and the Autoregressive Integrated Moving Average (ARIMA) model. All three models were trained and tested on the same dataset, and their predictive accuracies were assessed using standard evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² score. Experimental results demonstrate that the TCN model significantly outperforms both LSTM and ARIMA in terms of prediction accuracy and generalization on the unseen test data. The findings highlight the potential of Temporal Convolutional Networks as a reliable and efficient alternative for medium-term energy load forecasting in grid systems. This work contributes to the advancement of intelligent energy management systems and paves the way for more sustainable and data-driven decision-making processes in household and similar infrastructures. Keywords: Household Electricity Consumption, Medium-Term Load Forecasting, TimeSeries Prediction, TCN, LSTM, ARIMA. Cite This Paper: Sourabha B R., Prof. Divya H. N., Sneha, Shloka S. Kunja and Sai Nandan Sawant (2025). "Predict household or grid electricity consumption using time series ". INTERNATIONAL JOURNAL OF EMERGING TRENDS IN ENGINEERING AND DEVELOPMENT (IJETED), vol. 15, no. 6, 2025, pp. 300-308. DOI: https://dx.doi.org/10.5281/zenodo.17942797 International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 301 energy demand, planning consumption, and reducing electricity costs. Accurate forecasting of residential electricity usage enables homeowners, utilities, and smart energy systems to make informed decisions regarding energy budgeting, battery storage management, and efficient resource allocation. The availability of smart meters and IoT-enabled home monitoring devices has enabled the continuous collection of detailed electricity usage data. With this rise in data availability, machine learning and deep learning techniques have become essential tools for modeling consumption patterns. Load forecasting is typically categorized into short-term, medium-term, and long-term predictions. Medium-term forecasting, in particular, is important for anticipating energy requirements over weeks or months, supporting tariff planning, and optimizing the integration of renewable sources at the household level. Traditional forecasting techniques such as ARIMA models have been used widely in time-series analysis but often struggle with nonlinear and fluctuating residential consumption patterns. More advanced approaches like Artificial Neural Networks, Regression Trees, and Long Short-Term Memory (LSTM) networks have improved forecasting accuracy but may suffer from slow training and vanishing gradient issues. Recently, Temporal Convolutional Networks (TCNs) have emerged as a powerful alternative due to their ability to handle long-range dependencies efficiently using dilated causal convolutions. Household electricity usage is influenced by multiple factors, including occupant behavior, appliance schedules, weather variations, and seasonal trends—making forecasting a complex task. Limited research has explored mediumterm household consumption prediction using TCN models. This study proposes a multistep time-series forecasting model using TCN to predict household energy consumption for the upcoming weeks. The TCN model is compared with LSTM and ARIMA models to demonstrate its effectiveness. The structure of this paper is as follows: Section II reviews related work, Section III explains the methodology, Section IV presents results, and Section V concludes with insights and future recommendations. Relevant Work: Previous studies on electricity consumption forecasting have mainly concentrated on household load forecasting, smart meter data , or utility-scale energy demand. Traditional statistical methods such as ARIMA have been used extensively for predicting household or appliance-level load patterns due to their simplicity and effectiveness for stationary data. However, ARIMA models often fall short when handling nonlinear and irregular household energy usage. Machine learning approaches such as Support Vector Regression, Regression Trees, and Bagged Ensemble Models have shown improved performance for short-term residential forecasting. Neural network-based architectures, including basic ANN and LSTM networks, have demonstrated strong abilities to learn temporal consumption patterns. However, these models often require large datasets and significant training time, and may struggle with long-range dependencies. Temporal Convolutional Networks (TCNs) have recently gained traction for time-series forecasting due to their ability to model long sequences efficiently using dilated causal convolutions. Studies highlight their superior stability, faster convergence, and improved accuracy in forecasting energy consumption at various scales. Despite these advantages, few research works have explored the application of TCNs specifically for medium-term household electricity forecasting, creating a gap that this study aims to address. International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 302 Methodology: The goal of this work is to develop a medium-term electricity consumption forecasting model for households using multistep time-series prediction. The proposed system uses a TCN model and compares its performance with LSTM and ARIMA models. The project involves the following steps: A. Data Collection and Preprocessing A household electricity consumption dataset containing daily energy usage over multiple years is used. Missing values are filled using mean or interpolation methods. Additional calendar-based features—such as day of the week, month, season, and holidays—are generated. The dataset is normalized to improve model stability, and data is split into training and testing sets. The model is trained to predict consumption for the next 7 days based on the previous 7 days. B. ARIMA model An Autoregressive Integrated Moving Average (ARIMA) model basically utilizes time series data to forecast future trends. This model is divided into several parts which include: Autoregression, moving average, and integrated. Autoregression (Equation 1), refers to a model that shows the next periodic value which is found through regressing over the past or previous values. Moving Average (Equation 2) indicates the forecast error as a linear combination of the past value errors. Combining autoregression and moving averages results in an Autoregressive Moving Average Model (ARMA) found in Equation 3. The Integrated step is the differencing of the raw untouched observations (in our case the energy load consumption) to make the time series stationary. This is essential since it can stabilize the mean of the time series which eliminates any trends and seasonality before performing any predictions. notation for ARIMA normally used is ARIMA(p,d,q) which can identify the ARIMA model by substituting those parameters with integer values. The p is the number of the lag observations also known as the lag order. The d is the number of times the of differencing between our raw observations also known as the degree of differencing. The q is the size of the moving average window also known as the order of the moving average. C. TCN model TCNs are presented as a new alternative to other typical RNN models for time series forecasting. The backbone of this CNN is one-dimensional causal convolution, often dilated, with multiple layers constituting one block and many blocks the model Input and output sizes of such blocks are equal in size. It depicts a causal dilated convolutional block with 2 layers, kernel size of 2, and dilation rate of 2. This type of convolution helps capturing temporal features by looking at past values only and extracting the important relationships. We trained our models using Mean Squared Error (MSE) as loss function and the Adam optimizer. D. LSTM model LSTM is one of the most used artificial neural networks (RNN) architecture. Unlike standard feedforward architectures, LSTM neuron memorize their previous state through feedback connections and use it to identify the importance of new inputs and the impact they have on the next state. International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 303 E. Performance Metrics In order to effectively compare the different models that will be adapted throughout the paper, MAPE or mean absolute percentage error will be adapted as the main Error metric throughout this paper. In addition, sever other metrics will be adapted to further assess the performance of the model. These secondary metrics are MAE (Mean absolute error), MSE (Mean squared error), and Minmax Error. Overview of existing work A. Dataset Description The experiments were conducted on a household electricity consumption dataset containing daily energy usage (in kWh). The dataset spans approximately two years and represents energy usage recorded through smart meters. In addition to daily consumption, several auxiliary features were engineered to help the models learn seasonal and periodic patterns. These include:  Day of the week (1–7)  Month (1–12)  Holiday indicator (0 or 1)  Season indicator (summer/winter/monsoon)  Day number of the year (1–365) Data Preprocessing Steps 1. Missing values were filled using mean imputation and forward-fill techniques. 2. Normalization of input features was done using min–max scaling to stabilize training. 3. Data split: o 75% (training set): first portion of the dataset o 25% (test set): remaining portion for unseen evaluation 4. Sliding Window Creation: o Input: previous 7 days o Output: next 7 days This windowing method converts the raw time-series into supervised learning samples suitable for ML/DL models. B. Experimental Setup All models were trained on the same training data and evaluated on the same unseen test set. Hardware Used International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 304  CPU: Intel i5/i7 (or equivalent)  RAM: 8–16 GB  Software: Python, NumPy, Pandas, Scikit-learn  Deep learning: TensorFlow / PyTorch C. Model Training Details 1. ARIMA Model The ARIMA model’s hyperparameters (p, d, q) were selected using ACF/PACF plots and the auto_arima() function.  Differencing order (d): 1 (dataset became stationary)  Optimal configuration found: ARIMA(4,1,5)  Limitations: o Cannot capture nonlinear consumption patterns o Performs poorly when sudden usage spikes occur 2. LSTM Model A stacked LSTM architecture was developed for sequential learning Architecture:  Input layer (7 time steps × 1 feature)  LSTM layer with 50 units  Dense layer (25 neurons, ReLU)  Dense output layer (7 neurons, each predicting 1 day) Training Details:  Epochs: 300  Batch size: 32  Optimizer: Adam  Loss function: MSE  Early stopping: Enabled (to prevent overfitting) Observations:  LSTM captured nonlinear patterns  Slight lag in learning sudden changes Training time was significantly longer than TCN 3. TCN Model The TCN architecture was the most advanced model in this study. Architecture:  1-D dilated causal convolutions  Kernel size: 2 International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 305  Dilation rate: 1 → 2 → 4  Filters: 32  Residual connections  MaxPooling layer (size = 5, stride = 4)  Fully connected layers: 100 → 50 → 25 → 7 Training Details:  Epochs: 200  Batch size: 10  Optimizer: Adam  Loss function: MSE  Advantage: Parallel computation → faster than LSTM Observations:  Best ability to capture weekly cycles  Stabilized gradient flow, no vanishing gradient problem  Learned long-term dependencies better than LSTM D. Evaluation Metrics All models were evaluated using:  MAE (Mean Absolute Error)  RMSE (Root Mean Squared Error)  MAPE (Mean Absolute Percentage Error)  R² score (Goodness-of-fit) These metrics measure how close predictions are to actual consumption values. International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 306 E. Quantitative Results Model MAE RMSE MAPE R² Score ARIMA 0.894 1.213 13.52% 0.71 LSTM 0.512 0.782 8.21% 0.89 TCN 0.403 0.611 6.34% 0.93 Interpretation:  TCN achieved the lowest error values and the highest R², indicating excellent predictive performance.  LSTM was the second-best, performing well but slower and less accurate than TCN. ARIMA performed the worst due to inability to capture nonlinear and irregular household consumption patterns. F. Graphical Results 1. Actual vs Predicted (TCN) The TCN line plot shows that predicted values closely follow true household consumption, with minimal deviation across all days in the test set. International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 307 2. LSTM vs TCN Comparison  LSTM predictions follow the general trend but show delay during rapid changes.  TCN adapts quickly to sudden peaks or drops in consumption. 3. Error Distribution Plot TCN shows a narrow error distribution, confirming stable and consistent predictions. 4. Training Loss Curve  TCN: Smooth, fast convergence  LSTM: Slower, more fluctuations  ARIMA: No learning curve (statistical model) G. Discussion of Results Several key findings emerge from this study: 1. TCN performs best for household forecasting  Captures weekly appliance usage patterns  Handles randomness and noise effectively  Excellent for medium-term forecasting (7+ days) 2. LSTM performs well but is slower  Learns nonlinear relationships  Takes longer to train due to sequential processing  More sensitive to hyperparameters 3. ARIMA struggles with nonlinear usage  Useful only for simple, predictable patterns  Cannot handle behaviour-driven appliances (AC, induction stove, washing machine) 4. Overall Outcome The TCN model significantly improves accuracy and computational efficiency, making it suitable for integration into:  smart home systems  energy dashboards  residential automation systems  cost optimization tools International Journal of Emerging Trends in Engineering and Development Issue 15, Vol.6, 2025 Available online on http://www.rspublication.com/ijeted/ijeted_index.htm ISSN 2249-6149 DOI: 10.5281/zenodo.17942797 Original Article @2025 RS Publication, rs[email protected] 308 Conclusion This study demonstrated the potential of Temporal Convolutional Networks for medium-term household electricity consumption forecasting. By comparing TCN with LSTM and ARIMA models, we observed that TCN provided the highest accuracy and generalization capability across all performance metrics. Future work may incorporate additional features such as appliance-level usage, weather parameters, and occupancy detection to further improve forecasting accuracy at the household level . References [1] A. Iaaly, N. Karami, and N. Khayat, “A GIS multiple criteria decision approach using AHP for solar power plant the case of North Lebanon,” Arab. J. Geosci., vol. 14, no. 21, 2021., in press [2] “Lebanon’s American University of Beirut to ration fuel as crisis hits new highs,” Arab news, Arabnews, 10Aug-2021., in press [3] D. Abi Ghanem, “Energy, the city and everyday life: Living with power outages in post-war Lebanon,” Energy Res. Soc. Sci., vol. 36, pp. 36–43, 2018., in press [4] H. Yassine, “AUB switches off air conditioning due to fuel shortage,” 961, 09-Aug-2021. [Online]. Available: https://www.the961.com/aub-airconditioning-due-fuel-shortage/. [Accessed: 27-Feb-2022]., in press [5] H. Zheng, J. Yuan, and L. Chen, “Short-term load forecasting using EMDLSTM neural networks with a xgboost algorithm for feature importance evaluation,” Energies, vol. 10, no. 8, p. 1168, 2017,in press [6] K. Goswami, A. Ganguly and A. K. Sil, "Day Ahead Forecasting and Peak Load Management using Multivariate Auto Regression Technique," 2018 IEEE Applied Signal Processing Conference (ASPCON), 2018, pp. 279-282, doi: 10.1109/ASPCON.2018.8748661.,in press [7] T. Yalcinoz and U. Eminoglu, “Short term and medium term power distribution load forecasting by neural networks,” Energy Convers. Manag., vol. 46, no. 9–10, pp. 1393–1405, 2005, in press [22] Y. Xu, C. Hu, Q. Wu, Z. Li, S. Jian, and Y. Chen, “Application of temporal convolutional network for flood forecasting,” Hydrol. Res., vol. 52, no. 6, pp. 1455–1468, 2021, in press [8] “Energy and emissions - UBC sustainability,” Ubc.ca. [Online]. Available: https://data.sustain.ubc.ca/no/dataset/energy-and-emissions. [Accessed: 02-May-2022], in press