A Combinatorial Optimization and Analysis for Deep Learning Models of Short-Term Load Forecasting in Smart Grid

Power load forecasting is one of the important research topics in the field of power systems. Accurate power load forecasting can effectively reduce energy production costs and contribute to the sustainable development of humans and nature. In order to better dig out the effective information contained in massive data, four machine learning models for power load prediction were constructed in this study, which are LSTM, Conv1D-LSTM, Conv3D-LSTM and Adaboost. The electricity production, consumption and price of Spain from 2016 to 2020 were used as experimental data. Through training and experiment, the prediction effects, advantages and disadvantages of different models were analyzed, the RMSE and MAPE were used as the main evaluation index, and the MAE and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{R}^{\mathrm{2}}$</tex> determination coefficient were used as the auxiliary evaluation indexes. It is shown that the Conv3D-LSTM has the highest prediction accuracy under the same evaluation index, LSTM and Adaboost models have advantages when using data sets with small data volume and low attribute features, and the CNN-LSTM model has better universality, which was less dependent on selection and correlation analysis of feature.