Smart Grid Load Forecasting and Optimization Based on Deep Learning

A novel load prediction model, which combines VAE and GCN, is presented. Firstly, VAE is applied to study the potential features of power grid load data, which can effectively capture non-linear patterns in data. Secondly, GCN is used to model power network topology, and the spatial dependence of power network nodes in different areas is studied. Through the combination of VAE and GCN, the model is able to study the internal patterns of time series data and the interaction among them. To improve the training stability, a modified KL divergence loss function is proposed to solve the problem of unstable VAE in training. Then, the real load data of an intelligent grid is verified and compared with the conventional ARIMA, SVR and LSTM. The experiment results show that the VAE-GCN hybrid model is better than the traditional methods in MAE, RMSE, and MAPE, especially in long term load forecasting. At last, the paper designs an optimal scheduling strategy for grid load, which is based on deep reinforcement learning, which can efficiently allocate and adjust grid resources dynamically. This paper offers a new technology scheme to forecast and optimize the scheduling of the intelligent grid, and it is of great practical value.