Short-term Probabilistic Electric Load Forecasting Based On Deep Neural Networks

Electric load forecasting is crucial to the reliable operation of power system,and plays an essential role in energy management,economic dispatching and spot trading in smart grids.With the large-scale integration of renewable energy power generation and deregulation deepen in the power industry,the uncertainty of the power system will increase greatly.The traditional deterministic electric load point forecasting method will be not able to satisfy the requirements of the development of smart grids.However,the probabilistic load forecasting can successfully quantify the uncertainty of power demand and has received great attention from researchers.In this thesis,three new short-term electric load point forecasting models are proposed based on massive high-resolution smart meter data and deep neural networks(DNN).Also,two new short-term electric load probabilistic forecasting methods are proposed on the basis of the proposed point forecasting models.The main research work in this thesis is as follows:(1)The techniques of missing value filling and outlier correction are applied to the actual electric load dataset.Based on the characteristics of the short-term electric load,the empirical mode decomposition(EMD)algorithm is used to decompose the original load sequence.The load decomposition components are converted into a two-dimensional matrix as the input of a convolutional neural network(CNN),thus effectively helping the model to learn the local hidden features of the load time series at different time scales.On the basis of a similar day selection algorithm,loads of similar days are selected as the input of the point forecasting models and the probabilistic forecasting methods to provide additional effective features.(2)To deal with the feature extraction problem existing in the short-term load point forecasting,three short-term load point forecasting models based on multimodal DNN are proposed in this thesis.More specifically,these proposed short-term load point forecasting models use the VGGNet sub-network,Inception sub-network and Res Net sub-network to extract the spatial features hidden in the two-dimensional load EMD component matrix.Then,the spatial features are fused with load,electricity price and time information as the input of the long short-term memory(LSTM)sub-network.Finally,the LSTM sub-network captures the long-time dependence among the input data to estimate the load value for the next hour.Therefore,the three proposed forecasting models can extract multimodal spatial-temporal features,which contain more hidden information.The experimental results show that these proposed short-term load point forecasting models have higher forecasting accuracy than the traditional models.(3)To deal with the problem that the point forecasting is difficult to quantify the uncertainty of the electric load,a short-term load probabilistic forecasting method based on quantile regression random forests and a short-term load probabilistic forecasting method based on conditional residual modeling are proposed from the model side and the output side,respectively.Specifically,the former method uses the similar day selection algorithm and the three proposed point forecasting models based on multimodal DNN to obtain the transition point forecasting results.Based on the transition point forecasting results,the quantile regression random forests are applied to perform the short-term electric load probabilistic forecasting.The short-term load probabilistic forecasting method based on conditional residual modeling first obtains the transition point forecasting results by the three proposed point forecasting models and the similar day selection algorithm.Then,the random forests model is used to performs the point forecasting.Finally,the quantile regression random forests model the conditional residual distribution based on the point forecasting results and the transition forecasting results to achieve the short-term electric load probabilistic forecasting.The experimental results show that the two proposed probabilistic forecasting methods have higher short-term load probabilistic forecasting accuracy than the traditional methods.