Deep Learning Based Short-term Power Load Probabilistic Forecasting

Accurate load forecasting is the premise of maintaining real-time power supply and demand balance and ensuring the reliability and economy of power system operation.With the gradual intelligentization and cleaning of the power system,a large number of active loads and distributed renewable energy are connected to the power system,which increases the uncertainty of the load.Traditional point forecasting methods have been unable to meet the development needs of the power system.Probabilistic forecasting can forecast the complete probability distribution of the load and obtain more comprehensive load uncertainty information,which is one of the key means to solve the above problems.Deep learning has powerful feature extraction capabilities and is widely used in load forecasting.In this thesis,short-term probabilistic load forecasting based on deep learning is carried out for two highuncertainty scenarios of small-scale residential user load and the net load of the behind-themeter system with a large number of distributed photovoltaics installed.For the short-term power load probabilistic forecasting of small-scale residential users,this thesis proposes a probabilistic load forecasting model that combines variational modal decomposition（VMD）,temporal convolutional network（TCN）and quantile regression（QR）.The model first uses the VMD to decompose the load into multiple mode components to reduce the complexity of the load;considering the powerful time series processing capability of TCN and its characteristics of parallel computing and gradient stability,TCN is used as a deep learning load forecasting model.Load quantile forecasting results are generated directly using QR,and finally probability density distributions are generated using kernel density estimation.By comparing the load forecasting experiments with two probability generation methods based on point forecasting residuals,the high accuracy and efficiency of QR directly generating load quantiles are verified.The better forecasting performance of the TCN model in the load forecasting task is verified by the experimental comparison with the long shortterm memory network and the multi-layer perceptron.Because the distributed photovoltaic system is installed behind the meter in the behindthe-meter system,its output is invisible,which brings challenges to the forecasting of net load.Aiming at the probabilistic forecasting of the net load of the behind-the-meter system with a large number of distributed photovoltaics installed,this thesis proposes an indirect probabilistic forecasting method of the net load of the behind-the-meter system based on the photovoltaic decomposition technology.The method firstly uses photovoltaic decomposition technology to decouple the photovoltaic power generation data in the net load,obtains the estimated photovoltaic power generation data,and reduces the complexity of the net load.Then build a photovoltaic point forecasting model to forecast photovoltaic power generation data,and then use the above-mentioned combined VMD,TCN and QR model as a forecasting model to build a user load probabilistic forecasting model to model the uncertainty of user load and photovoltaic point forecasting residuals.The net load probabilistic forecasting result is obtained indirectly through the integration of the forecasting results of the photovoltaic point forecasting model and the user load probabilistic forecasting model.Numerical example analysis verifies that the proposed indirect forecasting method is better than the direct forecasting method in accuracy.