Residential Load Forecasting Based on CNN-LSTM and Non-uniform Quantization

Abstract

Accurate residential load forecasting plays an important role to improve the economy and security of power system operation. However, as the unbalanced distribution of residential load and the intertwined effects of multiple factors, it is difficult for a single neural network to make accurate predictions and its ability to generalize is limited. In this regard, this paper proposes a CNN-LSTM and non-uniform quantization based method for one-hour ahead residential load forecasting. First, we solve the unbalanced distribution of residential load by non-uniform quantization, which converts the load to an approximately normal distribution and fits the learning of neural networks. Then, the equivalent load after non-uniform quantization and its influencing factors are interwoven to form intertwining diagrams to facilitate the extraction of nonlinear relationships. Next, considering the intertwined effects of multiple factors, we use CNN-LSTM to extract temporal and spatial characteristics between multiple factors and cope with complex load patterns. We train and validate the proposed method using a real-world dataset, and the experiment results show that the proposed method outperforms the existing load forecasting methods.