The Research On Wind Power Forecasting Based On Deep Learning

Wind power plays an important role in the optimization of the global energy sector.Vigorously developing wind power generation is one of the important means to achieve carbon peak and carbon neutrality.However,with the increasing of wind power installed capacity,the fluctuation and uncertainty of wind bring huge risks and challenges to the safe and steady operation of power system.Therefore,it is particularly important to accurately forecast wind power.The traditional wind power forecasting method is still insufficient in dealing with massive data and nonlinear mapping ability.In recent years,the artificial intelligence technology is making great changes in many industries including finance,medical treatment and energy,which shows that using a new generation of artificial intelligence technology with deep learning algorithm as the core to solve the problem of wind power forecasting has a technical basis and industry needs.In this thesis,the wind power prediction model based on deep learning is proposed,which could provide decision support for the stable operation of the power grid.The uncertainty prediction of ultra-short-term wind power is discussed.The probabilistic prediction methods based on QRLSTM and QRDCC are proposed respectively.Adam algorithm is applied to estimate the parameters of the neural network under different quantile conditions.The predicted value of wind power under different quantiles at each moment in the next 200 hours was obtained by rolling prediction 1 hour in advance.Then,the probability density distribution of the predicted value at each moment was obtained by kernel density estimation method,which realized the prediction of the probability interval of wind power in the future.Finally,the wind power data from PJM network in the United States are used for example simulation.The results show that the probability prediction method proposed in this thesis can get a narrower interval average width under the premise of ensuring the interval coverage probability.The uncertainty prediction method based on QRLSTM and QRDCC models has a satisfying effect when applied to the ultra-short-term wind power prediction.However,in the long time scales of wind power forecast,the time invariance of the above models can reduce the capacity to perform multi-step-ahead forecasting.To address this,the time-variant deep feed-forward neural network architecture(ForecastNet)is used for short-term wind power probability prediction.Its architecture breaks from convention of structuring a model around the RNN or CNN.And the time-invariant problem of constructing model based on cyclic neural network or convolutional neural network is solved.The network structure of the model changes with time to improve the ability of multi-step-ahead prediction,the model is interleaved outputs to alleviate the problem of gradient disappearance,and the probability density distribution at each moment is obtained by using the mixed density network.This model not only avoids the cumulative error of recursive multi-step prediction in the traditional deep learning model,but also fully considers the correlation of wind power at adjacent moments.The thesis uses the actual data of wind power from PJM network in the United States.In the hidden layer of the model,three kinds of neural network models,namely,fully connected network,convolutional network,attention mechanism and convolutional network,are applied respectively to predict.The wind power of the next 12 hours is predicted each time,and the range and probability density of wind power of the next 500 hours are obtained by rolling prediction,the experimental simulation results prove the effectiveness of the proposed prediction model.In this thesis,the wind power uncertainty prediction model based on deep learning is proposed,which can provide accurate wind power fluctuation range for power system to make unit commitment and robust dispatching plan,improve the supporting value of wind power prediction in dispatching strategy optimization,and enhance the practical level of power prediction.