Hybrid Model for Long-Term Electricity Consumption Forecasting: Enhancing Accuracy with ElasticNetCV

Abstract

Reliable long-term electricity consumption fore-casting is essential for effective energy management, resource planning, and decision-making within the energy sector. This paper undertakes a comparative investigation into time series forecasting models for long-term electricity demand, evaluating established machine learning and statistical techniques alongside a novel hybrid ensemble approach. We assess various models, including XGBoost, Prophet, NeuralProphet, LSTM, GRU, and SARIMAX, using both historical and synthetically generated datasets. Our proposed hybrid ensemble model, which utilizes ElasticNetCV, synthesizes the strengths of individual models to improve accuracy, especially for extended forecasting horizons. Experimental results indicate that while deep learning models like LSTM and GRU perform strongly on historical data, our hybrid ensemble demonstrates superior relative performance on synthetic datasets crafted to simulate long-term scenarios. This underscores the enhanced adaptability and robustness offered by the ensemble methodology. We detail the specifics of data preprocessing, feature engineering, model configuration, training procedures, and performance assessment using standard error metrics. The findings highlight the potential for hybrid ensemble models to advance long-term electricity consumption forecasting, presenting a valuable direction for future research and practical applications.