Hybrid feature-based neural network regression method for load profiles forecasting

Abstract

This study addresses the critical need for improved demand forecasting models that can accurately predict energy consumption, particularly in the context of varying geographical and climatic conditions. The work introduces a novel demand forecasting model that integrates clustering techniques and feature engineering into neural network regression, with a specific focus on incorporating correlations with air temperature. Evaluation of the model’s efficacy utilized a benchmark dataset from Tetouan, Morocco, where existing forecasting methods yielded RMSE values ranging from 6429 to 10,220 [MWh]. In contrast, the proposed approach achieved a significantly lower RMSE of 5168, indicating its superiority. Subsequent application of the model to forecast demand in Astana, Kazakhstan, as a case study, showcased its efficacy further. Comparative analysis against a baseline neural network method revealed a notable improvement, with the proposed model exhibiting a MAPE of 5.19% compared to the baseline’s 17.36%. These findings highlight the potential of the proposed approach to enhance demand forecasting accuracy, particularly across diverse geographical contexts, by leveraging climate-related inputs, the methodology also demonstrates potential for broader applications, such as flood forecasting, agricultural yield prediction, or water resource management.