Edge-side load prediction analysis method based on hardware acceleration

The data for load forecasting is characterized by randomness and burstiness, and it is often required to have an accurate and fast load forecasting response in integrated energy and power market trading. Therefore, a solution is proposed for this scenario, which is based on an edge computing approach using an algorithmic model combining variable modal decomposition (VMD) and long and short-term memory network (LSTM). Because of the use of VMD, it is necessary to train multiple LSTM networks and complete the forward inference as simultaneously as possible, and the parallelization feature of Field Programmable Gate Array (FPGA) is considered for hardware acceleration design. First, edge computing has low latency, fast response, etc., and the data is decomposed into multiple intrinsic modal functions using variable modal decomposition so that the LSTM models can be trained separately to better handle nonlinear and nonstationary data. This is used to achieve computational acceleration and hardware resource consumption optimization. The results of the comparative analysis show that the edge-based VMD+LSTM model hardware architecture exhibits excellent performance in terms of computational speed for load prediction.