Two-Stage Ultra-Short-Term Load Forecasting Model of Household Appliances Based on Non-Intrusive Load Disaggregation

Abstract

Due to the impact of large-scale renewable energy on the safe and stable operation of power systems, the demand for flexible sources is increasing. The home energy management system is a promising approach to enhance the flexible regulation capability of power systems and improve grid energy efficiency. However, the randomness of residents' electricity behavior, the uncertainty of market information, and the diversity of decision-making subjects make it extremely challenging for residents to participate in fast demand responses. To address these issues, this paper proposes a two-stage household load forecasting method based on the integration of load disaggregation and forecasting. By learning the correlation information from historical electricity consumption data of each appliance obtained by non-intrusive load monitoring (NILM) technology, it accurately realizes the load forecasting of household appliances and flexible cluster load prediction. First, a NILM model based on convolutional neural network (CNN) and bi-directional gated unit (BiGRU) neural network is established to solve the problem of obtaining the operation data of appliances. Subsequently, considering the randomness and uncertainty of user behavior, a time convolutional network (TCN) load forecasting model based on the time pattern attention (TPA) mechanism is constructed to mine the deep interaction information of input variables. Finally, the proposed method is verified by the UK-DALE data set. The results show that the proposed method can obtain high disaggregation accuracy and prediction effect. This paper implements two simulations using Keras with a TensorFlow backend. The first one is designed to monitor the appliance-level energy consumption with the proposed CNN-BiGRU-enabled NILM. The results show that the proposed NILM-Based model can accurately capture the start and end time of the appliance, and has a good trend-tracking effect. With the NILM results, the second one is conducted to verify the effectiveness of the proposed load forecasting model. Compared with other deep learning models such as long short-term memory, the e_MAE and e_RMSE of washing machines, microwave ovens, dishwashers and refrigerators based on the proposed forecasting method are reduced by 18.65% and 4.99%, 13.28% and 0.72%, 32.04% and 5.55%, 4.53% and 5.70%, respectively. The comparison of load forecasting results between the ground truth data and the NILM-based data shows that the e_MAE is controlled within 15% and the e_RMSE is controlled within 6%. Finally, to verify the robustness of the proposed model in cluster load forecasting, a bottom-up strategy is applied to obtain the flexible load prediction of group users or communities based on appliance prediction. The following conclusions can be drawn from the simulation analysis: (1) The load forecasting framework based on NILM reduces the dependence on intrusive monitoring systems, and has strong adaptability and scalability, which can provide a new method for the development of household energy management system-related businesses. (2) Compared with shallow neural networks, CNN-BiGRU has powerful mapping ability and can achieve high accuracy in load disaggregation and event detection. (3) The proposed forecasting model extracts temporal informative features using TCN and subsequently implements dynamic weighting of inputs using TPA to highlight the impact of key features. The case study shows that the proposed method can improve forecasting accuracy compared with traditional deep learning models.