A hybrid model for non-stationary power load forecasting

In the context of China's electricity market, which is gradually promoting market-oriented reforms and establishing electricity spot trading, power load, as an important indicator of the electricity market, which can help industrial users adjust their production plans to avoid peak hours and save economic costs. However, load data is non-stationary and it is difficult for a single model to capture all the complex change patterns. Therefore, this study proposes a hybrid model based on trend decomposition and long short-term memory network to improve the accuracy of electricity load forecasting. The hybrid model divides the sequence into a period and a trend component by means of a double trend separation block. The non-stationarity of the periodic components obtained after trend decomposition is significantly reduced, making it easier for long short-term memory to learn intrinsic changes. Meanwhile, a simple multilayer perceptron is used to predict the trend component, and the final prediction is obtained by combining these two components. Experimental results show that the hybrid model performs well in different step-size forecasting tasks and outperforms other baseline models. The effectiveness of the trend decomposition module is also verified through visualization. This hybrid model can provide electricity market participants with more accurate load forecasts to support decision making and optimize resource allocation, which is of great practical importance.