Wind Power Forecasting Based On Deep Learning Algorithm

With the depletion of fossil fuels and their increasing pressure on the environment,the development and utilization of renewable energy is imminent.Wind energy is the fastest growing renewable energy in the world.Owing to the volatility,randomness and intermittency of the wind power itself,the phenomenon of wind curtailment is very serious.Wind power forecasting is an important reference for the site selection,large-scale grid connection and operation of the wind farm.With the development of numerical weather forecasting and big data technology,the traditional machine learning algorithms cannot meet the requirements of large-scale data training.However,the rise of artificial intelligence and deep learning algorithms provide a solid theoretical basis for wind power prediction technology.The combination of numerical weather forecast data and deep learning algorithm in the domain of space-time is a prospective research direction.At present,most research directions focus on how to apply deep learning algorithm to wind speed and wind power prediction.There is no systematic research and attempt to apply the data modeling of numerical weather forecasting and the application characteristics of various deep neural networks.Firstly,aiming at the characteristics of numerical weather forecasting,a variety of data modeling methods are proposed for the input of deep neural network.Based on the actual application requirements of the project,different data modeling methods and various deep neural networks were tested for short-term wind power prediction and ultra-short-term wind power prediction respectively.The prediction error characteristics and the causes of errors were analyzed.This paper mainly includes the following content:(1)Based on the multi-dimensional space-time characteristics of numerical weather forecast data,this paper proposed a variety of multi-dimensional spatial-temporal data structures suitable for deep neural network input,providing the basis for short-term wind power forecasting and ultra-short-term wind power forecasting input.(2)Based on the established multi-dimensional spatio-temporal data model,a variety of convolutional neural networks were used to make short-term wind power forecasting.Based on training and testing of actual wind field data,the feasibility of its superiority to traditional machine learning methods was verified,characteristics of deep neural networks with different structures were analyzed and single-site and multi-site predictions multiple NWP input were attempted.(3)Applying deep neural network architecture to improve the ultra-short-term power forecasting technology,designing an ultra-short-term prediction model based on multiple recurrent neural networks,and introducing Keras-based deep neural network construction and training methods.Based on the analysis of examples,deep learning algorithm's practicality and superiority in ultra-short-term wind power forecasting is illustrated.