Time Series Oriented Deep Neural Network For Load Probabilistic Forecasting

With the increasing proportion of new energy installations and the development of industrial structure,the power system has become increasingly large and complex.The traditional power system in the face of diversified demand and increasingly complex grid structure has generated problems such as supply-demand imbalance,uneconomic dispatch,low energy utilization efficiency,unreasonable expansion planning,etc.,which seriously affect the reliable operation of the power system and hinder the improvement of power economic efficiency,and electricity load forecasting technology is regarded as an important support to solve these problems.Electricity load data usually has certain periodic statistical characteristics,but due to the large uncertainty of user behavior,it leads to rich and variable characteristics of user load data,and the correlation and regularity are obviously weaker than the load data at regional level or bus level.Even if certain periodic statistical characteristics exist,the localization of load sequences still shows significant randomness.This leads to a significant increase in the difficulty of user-side load forecasting,and traditional load forecasting techniques are difficult to apply.This topic focuses on load probabilistic forecasting methods and load multi-step forecasting methods.In the load probabilistic forecasting part,on the one hand,the load probabilistic forecasting method EMD-PTCN is proposed,which builds the load probabilistic forecasting model around Temporal Convolutional Network(TCN)to improve the computational efficiency while ensuring the reliability of the temporal probabilistic forecasting;the EMD algorithm is combined to The Pinball quantile function replaces the loss function in the traditional model to make the model achieve the function of probabilistic forecasting.On the other hand,a probabilistic load forecasting algorithm based on the time-domain convolutional neural network algorithm and Huber quantile algorithm is proposed to overcome the problem that the load forecasting model based on Pinball quantile function generates certain training oscillations in the training process in the absence of significant features in the load data;the uncertainty of the electricity load forecasting is quantified and analyzed,and the deterministic forecasting and probabilistic forecasting are combined by Weighted Average Multiple Loss Function(WAML)to form a probabilistic forecasting model considering the distribution of deterministic forecast residuals.Finally,a probabilistic load forecasting model with high computational efficiency,parallel computation,and accurate forecasting results is constructed for shortand medium-term loads.In the load multi-step forecasting part,an internal recursive neural network model is proposed,which designs the recursive process inside the neural network model and updates the relevant parameters through the back-propagation mechanism,so that the recursive structure has time-varying characteristics;and the multi-step load forecasting method is proposed based on this model,and for the problem that the predicted sequence lags relative to the actual sequence in multi-step forecasting,the dynamic time regularization algorithm(DTW)is adopted to measure the sequence similarity considering The DTW is used to measure the degree of sequence similarity under the consideration of lags.