Research On Power Load Probability Forecast Based On Time Series Memory Neural Network

Accurate power load forecasting provides an important basis for power system dispatch and planning.As the power system enters the era of big data,uncertain factors such as distributed energy and new loads are also increasing.How to mine the potential information of the massive data of power load and measure the uncertainty of the load is of vital significance.Accordingly,based on the difference in load power characteristics under different weather conditions,day types and time points,this paper takes into account the time series and uncertainty of load power,and introduces time series memory deep learning network and quantile regression methods.The time series probability forecast method of monthly load curve has been studied in depth.The main research contents include:(1)Combining the measured data of power load in a certain area,it analyzes in detail the periodic characteristics of power load changes over time,as well as the influence of temperature,rainfall,and daily types on load changes,laying the foundation for the subsequent establishment of predictive models and the determination of input variables.(2)Considering the non-linearity and time sequence of the load data,compare the effects of LSTM(Long-Short Term Memory),GRU(Gated Recurrent Unit),JANET(Just Another NETwork)and MGU(Minimum Gating Unit)on the deterministic prediction of day-ahead load curve by the four time-series memory neural networks,and further analyze the impact of changes in model parameters,training samples and input parameters on the prediction effect.The results show that: the four temporal memory neural networks can accurately predict the time series trend of load,but the prediction accuracy of LSTM and GRU is higher,and the training efficiency of JANET and MGU is higher.In addition,regarding the load time series forecasting problem and the sample size itself,there are a suitable network depth and number of neurons.Among them,the optimal time sequence length is closely related to the periodicity of the load change;while training on working days and rest days separately,and adding the previous day’s load data to the input parameters,the prediction accuracy and efficiency can be further improved.(3)Based on the research of day-ahead time series load deterministic forecasting,addressing the uncertainty of power load and the overlap between the adjacent quantile forecasts of the existing quantile regression methods,combined with LSTM neural network and quantile regression methods,this paper proposes a day-ahead power load curve probability prediction method based on quantile regression of constrained parallel LSTM neural network.This method can generate multiple quantile results of the load at each time point of the forecast day in parallel,and ensure the reasonableness of the quantile forecast value by adding a combination layer that considers the constraint relationship among the quantile forecast values.In addition,data parallel training methods are used for training to improve prediction efficiency.(4)Based on the probabilistic prediction method of the day-ahead load curve,in view of the problem of poor interpretability of LSTM,fixed length of the input and output sequence,and the inability to use the known information of the forecast day on a longer time scale in the future,a medium-term load probability forecasting method based on improved temporal fusion Transformer model is proposed,which extends the time scale of load probability forecasting from short-term day-ahead forecast to mid-term monthly forecast.Based on the daily periodicity of the load,the method reconstructs the original load time series into panel data,and inputs the time of day as the data label of each object in the cycle to learn the time series relationship of the load change at the same time and expand the time scale of time series.And an explanatory variable selection network is constructed to explain the specific contribution of each input variable on the output.In addition,the evaluation index reflecting the constraint relationship among the quantiles and the sharpness of the corresponding quantile prediction interval is added to the model loss function to ensure the reasonableness of the prediction results and construct a more compact prediction interval.The actual calculation example results show that the power load probability forecasting method based on the time series memory neural network proposed in this paper can better learn the dynamic change trend of the load and provide accurate probability forecast results,which can provide more accurate and rich information for the probability optimization dispatching,risk analysis and decision-making of the power system.