The Effect of Input Length on Prediction Accuracy in Short-Term Multi-Step Electricity Load Forecasting: A CNN-LSTM Approach

Abstract

Accurate load forecasting is crucial for effective power system management and planning in the context of growing electricity demand triggered by the proliferation of technological devices and rapid digitalization. Since electrical energy is largely non-storable, short-term electrical load forecasting plays a critical role for system operators. This paper presents an innovative hybrid deep learning model that combines convolutional neural networks (CNNs) and long short-term memory (LSTM) networks for short-term multi-step load forecasting using real-time hourly data from a residential customer. The model is tested on 12 different configurations with symmetrically increasing input lengths, including weather data. The results show that increasing the input length improves the learning performance of the model for all conditions. In addition, selecting an input length greater than the output length has been shown to improve prediction accuracy, with an improvement of 67% in Mean Absolute Percentage Error (MAPE) and 70% in Root Mean Square Error (RMSE). Moreover, it was observed that the multi-step forecasting performance with increased input length is more successful than the single-step forecasting performance.