An explainable and latent space-based deep learning model for improved short-term load forecasting

A rapid transition towards clean, green, and eco-friendly power systems necessitates the integration of renewable energy sources into the power grid. However, the intermittent nature of these renewable sources requires the integration of an accurate short-term load forecasting model to predict energy usage patterns and provide more reliable operation. To address this problem, this paper presents a novel framework comprising of: 1) an autoencoder to acquire more meaningful features from the raw dataset by mapping the dataset into latent space; 2) a deep learning framework to improve and automate the forecasting process based on convolutional and long short memory network techniques; and 3) an explainable artificial intelligence framework to comprehend and interpret the prediction generated by the model, and highlight the prominent features contributing to the underlying dataset. The proposed framework is evaluated using a comprehensive dataset encompassing power usage patterns over different time spans. The results are compared with conventional and state-of-art methods, which demonstrate that the proposed XAE-ConvLSTM achieves an overall improvement in mean absolute error of 189.80 ± 0.68% in AEP and 28.22 ± 0.08% in Panama, root mean squared error of 254.01 ± 0.67% in AEP and 38.60 ± 0.925% in Panama, mean absolute percentage error of 1.29 ± 0.071% in AEP and 2.32 ± 0.040% in Panama, and the coefficient of determination R ² 0.9892 ± 0.0028% in AEP and 0.9568 ± 0.0019% in Panama datasets. This research enhances the existing knowledge on predicting and explaining peak load demand and has practical implications for creating intelligent energy management systems in smart grid applications.