电工技术学报2024，Vol.39Issue(7)：2104-2115,12.DOI:10.19595/j.cnki.1000-6753.tces.230110
基于时空注意力机制的台区多用户短期负荷预测
Short-Term Load Forecasting for Multiple Customers in A Station Area Based on Spatial-Temporal Attention Mechanism
摘要
Abstract
With a large number of customer-side distributed power sources entering the network from low-voltage distribution stations and the widespread use of devices such as smart meters on the customer side,a load forecasting model for multiple users is needed to facilitate point forecasting and probabilistic tasks for a large number of users efficiently and accurately.Traditional methods of customer load forecasting model the temporal characteristics of individual customers and are unable to learn the problems of spatial correlation between customers and the inability to achieve forecasts for multiple customers.Customers in the same region share the same geographic space,weather conditions,holiday information,tariff policies,and other comprehensive factors,and there is often a certain amount of spatial and temporal correlation between customers'electricity consumption behavior.If this spatial-temporal correlation can be fully explored,it will have extremely positive implications for modeling short-term customer loads.A small body of literature has already explored the initial exploration of customer load forecasting,taking spatial-temporal correlation into account.However,the existing spatio-temporal methods can only provide deterministic forecasts,not probabilistic ones.To address these issues,this paper proposes a multi-customer short-term load forecasting model for station areas.Learning spatial-temporal correlation information from historical load data can perform accurate multi-user short-term load point forecasts and probabilistic forecasts for station areas. Firstly,three modules are embedded for each encoder and decoder by improving the standard Transformer self-attention mechanism:sequence decomposition module,autocorrelation calculation module,and spatial attention module to effectively extract the dynamic spatio-temporal dependencies among highly volatile residential users.Among them,the sequence decomposition module can decompose highly volatile subscriber load curves into relatively smooth multiple sub-series,which helps to extract better the time dependence and period factor of load curves;the autocorrelation calculation is an improved attention mechanism that can mine the time dependence of multiple historical contemporaneous sub-series;and the spatial attention mechanism can extract the dynamic spatial support among multiple users in a station area.The STformer model is then extended to the field of probabilistic forecasting using a Monte Carlo stochastic deactivation method(MC dropout).This method does not require additional modifications to STformer but allows STformer to output both point prediction and probabilistic prediction results.Finally,the STformer model with MC dropout is used to forecast the station customer load,and both point and probabilistic forecasts are output. In this paper,the model's validity is verified using one-hour-ahead load forecasting and day-ahead load forecasting using accurate station customer load data from a province in the southeast.The proposed STformer model has a MAPE of 4.44% for each user and 2.21% for the total load in station area A.The probabilistic forecast evaluation index pinball is 0.370 1;the average relative error MPE for each user and 3.25% for the total load in station area A is 6.21% .is 3.25% ,and the probabilistic forecast assessment index pinball is 0.594 2.This paper also compares the effects of different modules on the experimental results through ablation experiments.This paper also verifies the change in model inference speed brought about by the addition of FFT,comparing the running memory and time of the autocorrelation-based model with that of the self-attentive-based model during the training phase. The following conclusions can be drawn from the simulation analysis:(1)Compared with other baselines,the STformer model proposed in this paper extracts the temporal variation pattern of users through the temporal attention mechanism and the spatial dependency between multiple users through the spatial attention mechanism,which ultimately achieves the best prediction results in all scenarios.(2)Each module of STformer contributes to the improvement of prediction accuracy and model robustness.The spatial attention module has the greatest impact on the prediction accuracy of STformer,and the Fourier transform method of the autocorrelated model reduces the computational complexity and thus accelerates the computational speed of the model.(3)The prediction intervals of the proposed STformer model with MC dropout have reliable coverage of the true values and provide narrower prediction intervals,especially at some peaks and troughs,which are critical for the temperature operation of power systems.关键词
多用户负荷预测/时空相关性/Transformer模型Key words
Multi-customers load forecasting/spatial-temporal correlation/Transformer model分类
信息技术与安全科学引用本文复制引用
赵洪山,吴雨晨,温开云,孙承妍,薛阳..基于时空注意力机制的台区多用户短期负荷预测[J].电工技术学报,2024,39(7):2104-2115,12.基金项目
国家电网公司总部科技项目"基于智能量测的低压高渗透率分布式光伏接入可测可控技术研究"(5700-202255222A-1-1-ZN)资助. （5700-202255222A-1-1-ZN）
