Residential load forecasting using wavelet and collaborative representation transforms

•Appliance-level meters are used together with the whole data.•The lagged load values are used for feature extraction.•The details obtained by discrete wavelet transform are removed.•Collaborative representation integrates the load information related to past and future time instances. Short-term household-level load forecasting requires to acquire knowledge about lifestyle and consumption patterns of residents. A new forecasting framework is proposed in this work which uses the extra appliance measurements in meter-level for short-term electrical load forecasting. The long short-term memory network as a deep learning method is used as a predictor where useful features are fed to it for forecast learning. A lagged load variable vector is assigned to each point of the load curve. To remove redundant details and to use the approximate component of the feature vector, the wavelet decomposition is applied to it. In addition, a new version of collaborative representation is introduced and used to achieve information of the neighboring points (previous and future time instances) of the considered load point. Collaborative representation of the feature vector associated with each load point contains valuable local information about adjacent load points. The load features extracted from the lagged load variable vector provide superior forecasting performance especially with extra appliances load data.