Short Term Power Load Forecasting Method Based on VMD-ResNet

The load of a region is often influenced by multiple factors, resulting in strong nonlinearity and temporal characteristics of load data, which increases the difficulty of accurately predicting load. To improve the accuracy of power load forecasting, this paper proposes a short-term power load forecasting method based on VMD-ResNet. Firstly, at the data level, the Variational Modal Decomposition (VMD) method is used to decompose historical loads and temperatures to obtain a more stable Intrinsic Mode Function (IMF). Then, different IMFs are integrated and input into the first stage prediction model along with seasonal and holiday information to extract internal features of the data for preliminary prediction; The second stage is a Deep Residual Network (ResNet) that uses the forecast data from the first stage as input to predict the load over the next 24 hours. The data adopts real load data from Panama City from January 2015 to June 2020. The average absolute percentage error (MAPE) of the VMD-ResNet load forecasting model is 1.3%. Compared with load forecasting networks such as SVR, LSTM, CNN-LSTM, ResNet, etc., the MAPE of the test set has decreased by 2.36%, 1.8%, 1.2%, and 0.8%, respectively. Experiments have shown that VMD-ResNet can deeply mine data characteristics and has good predictive performance.