Abstract Regional integrated energy system (RIES) plays an important role in the energy economy because of its advantages such as low environmental pollution and high efficiency cascade energy utilization. In order to ensure the operational efficiency and reliability of RIES, the accurate prediction of energy demand has become a crucial task. To this end, this paper proposes a novel multi-energy load prediction model based on deep multi-task learning and ensemble approach for RIES. Its novelty lies in the following four aspects: (1) considering the high-dimensional temporal and spatial features, a hybrid network based on convolutional neural network (CNN) and gated recurrent unit (GRU) is utilized to extract high-dimensional abstract features and model nonlinear time series dynamically; (2) to meet the prediction requirements of various loads, three GRU networks with different structures are designed, which can adapt to different types of loads with various fluctuations; (3) considering the coupling relations, an enhanced multi-task learning with homoscedastic uncertainty (HUMTL) is proposed, which can better make the prediction tasks of various loads achieve the optimum simultaneously; (4) to realize the sharing of learning results of different structure networks, ensemble approach based on gradient boosting regressor tree (GBRT) is adopted, which can make a weighted summary by the prediction results of various energy features learning in different degrees. Numerical example shows that the proposed model can dig the coupling relations among various energy systems deeper, explore the temporal and spatial correlation of multi-energy loads further, and it has higher prediction accuracy and better prediction applicability than other current advanced models.

中文翻译：

---

基于深度多任务学习和集成方法的区域综合能源系统多能源负荷预测模型

摘要 区域综合能源系统（RIES）具有环境污染小、能源梯级利用效率高等优点，在能源经济中发挥着重要作用。为了保证 RIES 的运行效率和可靠性，对能源需求的准确预测已成为一项至关重要的任务。为此，本文针对 RIES 提出了一种基于深度多任务学习和集成方法的新型多能量负荷预测模型。其新颖之处在于以下四个方面：（1）考虑到高维时空特征，利用基于卷积神经网络（CNN）和门控循环单元（GRU）的混合网络提取高维抽象特征和动态建模非线性时间序列；(2)满足各种负荷的预测要求，设计了三种不同结构的GRU网络，可以适应不同类型的负载波动；(3)考虑耦合关系，提出了具有同方差不确定性的增强多任务学习(HUMTL)，可以更好地使各种负载的预测任务同时达到最优；(4)为了实现不同结构网络学习结果的共享，采用基于梯度提升回归树（GBRT）的集成方法，可以通过不同程度的各种能量特征学习的预测结果进行加权总结。数值算例表明，该模型可以更深入地挖掘各种能源系统之间的耦合关系，进一步探索多能源负荷的时空相关性，