基于电器运行状态和深度学习的非侵入式负荷分解

根据不同电器运行状态数的差异,将电器分为状态复杂和状态简单2类.状态复杂电器存在多种工作状态,且前后状态有逻辑关联.因此,利用非基于事件的方法,选择能考虑过去和未来运行状态变化的双向长短期记忆网络对其进行分解,并采用树结构Parzen估计算法选择该网络的超参数以提高训练的精度.状态简单电器仅有开关状态,故利用基于事件的方法获得其投切状态,并选择多层感知器网络识别对应电器种类.最后,利用极大似然优化模型求解电器的功率序列.利用参考能量分解数据集对所提方法进行验证,结果表明该方法增强了负荷分解模型的可拓展性和抗噪声能力,在一定程度上提高了负荷分解的精度.

Non-intrusive Load Decomposition Based on Operation State of Electrical Appliances and Deep Learning

According to the number of operation states of different electrical appliances, electrical appliances are divided into two types: complex-state and simple-state. Complex-state electrical appliances have multiple working states, and the prior and subsequent states are logically related. Therefore, a non-event-based method is adopted and the bidirectional long short-term memory network considering past and future operation state changes is chosen to decompose the load. In order to improve the training accuracy, the hyper-parameters of the network are chosen by using the tree-structured Parzen estimator algorithm. The simple-state electrical appliances only have on or off state, so the switching states are obtained by the event-based method. Then, the multilayer perceptron network is selected to identify the corresponding electrical appliance type. Finally, the maximum likelihood optimization model is used to solve the power sequence of the electrical appliances. The proposed method is verified by using the reference energy disaggregation data set. The results show that the method enhances the scalability and anti-noise ability of the load decomposition model, and improves the accuracy of the load decomposition to a certain extent.