Ultra-Short-Term Multistep Building Power Prediction Based on WD-BO-LSTM

Abstract

The power load of large public buildings has strong randomness, which is affected by many factors such as weather and personnel activities, so it is difficult to greatly improve the prediction accuracy by using traditional prediction methods. Besides, according to some site requirements, the single-step prediction could not provide sufficient forecast information. In order to solve the above problems, first, the building power sequences are decomposed using Wavelet Decomposition (WD). Then, Long Short-term Memory (LSTM) and Bayesian Optimization (BO) algorithm are used to build a prediction model for each component signal, and the influence of future weather is also considered in the model. In addition, the rolling multi-step prediction method, which is one of the multi-step prediction methods that will get more accurate information after time t than the one multi-step prediction, is selected according to the field requirements. Finally, we obtain 31 days' real power load data of public buildings through professional data acquisition system for prediction and analysis. The results show that the Root Mean Square Error (RMSE) of the proposed ultra-short-term multi-step prediction method is only 126.39. The accuracy of WD-BO-LSTM is 43.49% higher than that of the traditional LSTM network.