Short-Term Wind Speed Prediction Based On Deep Transfer Models

Owing to the depletion of conventional energy sources and the deterioration of the environment,clean and renewable energy sources are being widely utilized all around the world.With the advantages of non-pollution,low costs and large scale,wind power is considered as one of the most important sources of energy.However,as the most important factor of wind power,wind speed is uncertain and variable.Accurate prediction of wind speed is important for the allocation,scheduling,maintenance,and planning of wind energy conversion systems.As to some newly-built wind farms or wind turbines,sufficient historical data is not available for building an accurate model,while some older wind farms or turbines may have long-term wind speed records.Therefore,a question whether the prediction model can be built with the help of the data from other wind farms of turbines is considered.In this paper,we propose a solution of transferring the information obtained from data-rich farms of turbines to the target.The main contributions of the work are listed as follows:1.A new research question for wind speed prediction is introduced,i.e.,transferring knowledge from data-rich source domains to the target domain for its forecasting model.2.We analyze the variation characteristics of wind speeds and perform a statistic analysis on wind speed series of different lengths.And we compute the correlation coefficients between the target domain and every candidate sequence,then choose the more similar ones as the source domains.3.A transfer model based on shared-hidden-layer deep neural network is proposed.In this architecture,the hidden layers are shared across all domains and can extract shared abstract information.However,the output layers are separate from each other,then the every own variation patterns of wind speed can be better learned.4.A deep neural network transfer model considering temporal characteristic is proposed.The lower-level DNN is acted as a feature extractor,which learns shared high-level features with the whole historical data from all the domains.And input the transformed features belonging to each domain into the upper-level corresponding RNN.Therefore,the model combines the ability of feature selection with deep neural networks and consideration of temporal correlation with recurrent neural networks.The two proposed transfer models are applied to the wind speed prediction of the target domain which has a small volume of historical data.The experimental results show that our approachs outperform other algorithms and the latter model considering temporal characteristic performs better than the former approach.And we also demonstrate that the importance of the transferred knowledge decreases if more training data are available in the target domain.