A Dual-Model Approach to Residential Load Forecasting Us-ing CNN–BiLSTM and Multi-Step Temporal Architectures

Abstract

Introduction: Accurate forecasting of residential electricity consumption is crucial for optimizing energy management and ensuring sustainable power usage. This paper proposes comparisons two novel hybrid deep learning models tailored for short- and long-term electricity consumption forecasting: Hybrid CNN–BiLSTM Attenuation and Hybrid Temporal Fusion Transformer–CNN (HTFT-CNN) for accurate residential electricity consumption forecasting. The first model leverages Gabor filtering, Greedy Stepwise Correlation, and Random Forest-enhanced CNN–BiLSTM to capture short-term spatiotemporal patterns. The second combines CNN with a Temporal Fusion Transformer, employing attention mechanisms to provide reliable multivariate, multi-step forecasting. Based on experimental findings, CNN-BiLSTM model works better when making short-term predictions and HTFT-CNN works better in long-term prediction. Both outperform traditional methods, and they offer greater accuracy in the management of sustainable energy

Objectives: To develop two distinct predictive modelling algorithms to predict the residential electricity usage. and evaluate the performance of both proposed algorithms on a scale of time horizons and configurations of input features using common metrics, such as RMSE, MAPE and R2. The effectiveness of the various methods used to predict the forecasting results has to be investigated by comparing the quality of interpretability, computational costs and generalization abilities to unseen data of each technique in use.

Methods: In the analysis, the amount of variance in the prediction of electricity consumption using two emerging forecasting models (proposed Model 1 (CNN- BiLSTM) and proposed Model 2 (Multi-Step Forecasting)) is compared with traditional models with a view of deciding on their application in electricity consumption forecasting. It has been found that the predictive accuracy, robustness and computational performance of the two proposed model are much improved. CNN-BiLSTM model performs particularly well when modelling complex short-term variations as it makes use of convolutional layer to exploit spatial patterns and bidirectional LSTM layers to model both forward and backward complex temporal reliance’s. The Multi-Step Forecasting on the contrary, through systematically recording the temporal relationships over various time scales is believed to give accurate long-term predictions through the use of CNN-TFT model. Comparative analysis shows that Multi-Step Forecasting model is the more suitable one to calling tasks of the short-term decision requirement, and the CNN-BiLSTM model does it with great precision.

Results: The UCI Household Power Consumption dataset, which includes time-series energy consumption records from 2006 to 2010, was evaluated using 2,075,259 observations, sampled every minute. Data preprocessing involved handling missing values through forward filling and interpolation, applying Min-Max scaling for normalization, and splitting the dataset into 70% training, 15% validation, and 15% testing. In table 1, the proposed models are compared against existing machine learning and deep learning methods. The proposed models, proposed 1 (CNN-BiLSTM) and Proposed 2 (Multi-Step Forecasting), aimed to enhance forecasting accuracy by integrating CNN for feature extraction with BiLSTM for sequential learning and optimizing sequence length (k=4, 8, 12, 24) for multi-step forecasting. The proposed method achieves the least MAE (0.157), RMSE (0.238), and MAPE (4.02%), and highest R² value (0.951), denoting the best forecasting ability. The proposed 1 (CNN-BiLSTM) is also a strong contender and comes second in accuracy, whereas it produces lower MAE (0.165) and RMSE (0.249) than various traditional machine learning and standalone deep-learning models like LSTM, BiLSTM, and CNN. However, it lags just a little behind Proposed 2. RF and SVR performed poorly with the highest error and lowest R² score, confirming their inability to handle complex time-series dependencies.

Conclusions:

The assessment of the two novel frameworks developed and tested in this research study: Proposed 1 (CNN-BiLSTM) and Proposed 2 (Multi-Step Forecasting) against existing methods for forecasting electricity consumption. From the comparative study, it can be inferred that both proposed methods enhanced prediction X, robustness, and computational efficiency to a significant extent. CNN-BiLSTM learned spatial features through convolutional layers and modelled sequential dependencies through LSTM layers in both directions, thus being capable of capturing extremely sophisticated temporal changes. On the other hand, Multi-Step Forecasting CNN-TFT focuses on enhanced consistency in long-term predictions by working in a structured framework considering multi-scale temporal dependencies.