Fusion ConvLSTM-Net: Using Spatiotemporal Features to Increase Residential Load Forecast Horizon

Abstract

Power systems are undergoing a significant transition towards renewable energy technologies. To make the most of these energy sources, optimizing the generation, storage, and distribution of energy can be enhanced with accurate forecasts of future energy consumption. Forecasting the load of individual residents plays a key role in load balancing, but it remains challenging due to the irregular nature of individual consumption patterns. Moreover, the current literature is limited to forecasting residential load to only a few hours in the future. In this paper, we propose Fusion ConvLSTM-Net, a novel fusion encoder-decoder architecture that combines both spatial and temporal features to extend the load forecast to a full 24 hour period. We evaluated the model against several benchmark neural network models by: 1) testing different forecast window sizes ranging from 1.5 to 24 hours, 2) assessing model performance across multiple households, and 3) performing large-scale forecasting by aggregating predictions from 100 households. Additionally, we analyzed the model’s forecasts to identify potential degradation. Our extensive experiments demonstrate that the Fusion ConvLSTM-Net not only extends the forecast window to 24 hours but also reduces the prediction error rate by approximately 47% compared to the next best model, improves the accuracy of aggregate load forecasts, and prevents model degradation.

Keywords

• Demand forecasting
• short-term load forecasting
• artificial neural networks
• deep learning
• residential load forecasting