基于时序特征灰度图与多任务学习的综合能源负荷短期预测

Short-term Prediction of Integrated Energy Load Based on Time Series Feature Grayscale Map and Multi-task Learning

准确预测综合能源系统(Integrated energy system,IES)中的电、冷、热多元负荷是提高各类能源综合效率、获得更大经济效益的关键。因此,提出一种基于时序特征灰度图与多任务学习的综合能源负荷短期预测方法。首先将初始特征集通过最大互信息系数(Maximum information coefficient,MIC)改进的快速相关滤波算法(Fast correlation-based filter,FCBF)对IES时序特征数据集进行相关性分析和冗余性分析;然后将特征选择结果利用因数重构法与MIC-gamma图像增强的方法重构为时序特征灰度图,能够直观有效地反映实际数据的特征相关性;其次采用基于多任务学习框架的(Convolutional block attention module-convolutional neural network-deep bidirectional gated recurrent unit,CBAM-CNN-DBiGRU)网络进行训练,嵌入的卷积注意力机制模块(Convolutional block attention module,CBAM)与(Deep bidirectional gated recurrent unit,DBiGRU)结构能有效加强共享层的关键信息提取和时序信息处理能力;最后以美国亚利桑那州立大学的IES数据为例对提出的方法进行测试。选取典型工作日和典型休息日并对比多种深度网络模型,测试结果表明,该模型在典型工作日的加权平均绝对百分比误差与加权均方根误差分别最大降低了0.8813%与229.2593 kW,在典型休息日则分别最大降低了0.9942%与360.8007 kW,能够有效提升IES多元负荷预测精度。

Accurately predicting the multi-energy load of electrical,cooling and heating in the integrated energy system(IES)is the key to improving the comprehensive efficiency of various energy sources and obtaining greater economic benefits.Therefore,a method for short-term prediction of integrated energy load based on time series feature grayscale map and multi-task learning is proposed.Firstly,the correlation analysis and redundancy analysis of the IES time series feature data set are carried out through the initial feature set through the fast correlation-based filter(FCBF)improved by the maximum information coefficient(MIC).Secondly,the feature selection result is reconstructed into a time series feature grayscale image using the factor reconstruction method and the MIC-gamma image enhancement method,which can intuitively and effectively reflect the feature correlation of the actual data.Then,the CBAM-CNN-DBiGRU network based on the multi-task learning framework is used for training.The embedded convolutional block attention module(CBAM)and the deep bidirectional gated recurrent unit(DBiGRU)structure can effectively strengthen the key information extraction and time series information processing capabilities of the shared layer.Finally,the method proposed in the paper is tested by taking the IES data of Arizona State University as an example.Selecting typical working days and typical rest days and comparing various deep network models,the test results show that the weighted average absolute percentage error and weighted root mean square error of the model on typical working day is respectively reduced by a maximum of 0.8813%and 229.2593 kW,and on typical rest day is respectively reduced by a maximum of 0.9942%and 360.8007 kW,which can effectively improve the accuracy of IES multi-energy load prediction.