A Multi-Frequency Memory Network for Short-Term Electricity Load Forecasting

In this study, a novel forecasting model named multi-frequency memory network (mFMN) is developed for short-term power load forecasting to cope with the forecasting challenges caused by multiple complex factors, aiming to improve the reliability of power system scheduling. The mFMN first utilizes a multilevel discrete wavelet decomposition network (mWDN) to perform in-depth time-frequency profiling of power load data, revealing the hidden multiscale and multi-frequency features of the data, including the periodicity, trend, and transient changes of load fluctuations. Next, the model is equipped with multiple independent memory network modules corresponding to different frequency levels to capture and model the nonlinear dynamics and long-term historical correlations in load changes. The long-range memory property of the memory network ensures effective identification and storage of complex time series dependencies. The fully-connected network, as the final link of the model, integrates the comprehensive features from each frequency level and generates the power load forecasts for the next 24 hours through weight optimization. mFMN's core innovation lies in the introduction of a wavelet decomposition mechanism with adjustable parameters, which endows the model with higher adaptive and generalization capabilities, enabling it to better cope with the complex variability of power load data while maintaining the forecasting accuracy. Through rigorous comparison experiments on real power grid datasets, the paper verifies the effectiveness and advantages of the mFMN model.