A Short-term Load Forecasting Method Combining Multi-scale Analysis With Data Co-transfer

With the development of human society,people's demand for electric energy is also rising.However,some particular properties of electric energy have raised a rigid need to the construction and maintenance of power system.The power system also has an important impact on economic stability and social security.Short-term power load forecast is related to many aspects of power system.It is also an important part of the construction of smart grid.Efficient and accurate short-term power load forecasting makes the power system operate safely and stable,It can also increase economic efficiency,let people make plan for coal combustion more reasonably to promote ecological environment protection.This paper proposes a short-term load forecasting method combining multi-scale analysis with data co-transfer.In order to improve forecasting performance,this paper starts with improving two following aspects:Aiming at the problem that the hidden information in the original series which is related to the subseries did not fully utilized in the modeling and prediction of subseries in multi-scale analysis forecasting method.We apply mutual information feature selection method to select appropriate past loads and introduce them into the feature set of the approximation component of original load series.By expanding feature set,we can provide more information to the learning algorithm,which can improve the forecasting accuracy of the approximate component.Aiming at the problem that the difference of seven types of data corresponding to 7 days of one week can influence model's performance.We proposes a transfer learning framework based on kernel ridge regression which uses an instance-based transfer learning method to transfer the hidden knowledge from data of other day types to data corresponding to days to be forecasted.By transferring knowledge among data of different types of day,the learning framework can not only take advantage of the similarity of these data,but also take into account the difference of these data during modeling.Experiments show that the method proposed in this paper outperforms in MAPE,MAE and RMSE which is decreased by 6.2%,3.4%and 5.5%respectively when comparing with single model forecasting method.