Study On Renewable Energy Source-load Prediction Based On Deep Learning Neural Networks

The development of economy and society is inseparable from the large-scale consumption of fossil energy.The increasingly growing carbon emissions have led to increasingly severe environmental problems.The refore,the development of renewable energy has received great attention from countries all over the world.The impact of renewable energy on energy system reliability cannot be ignored.Predicting the output of renewable energy and user loads can help to respond timely to changes in supply and demand,optimize energy scheduling,or optimize system planning and design,thereby achieving a more economical,efficient,and low-carbon energy utilization mode.Statistical analysis models and some classical machi ne learning algorithms are commonly used in various prediction tasks.However,as energy systems become increasingly complex,these traditional methods may not be well suited to handle complex data scenarios.Emerging artificial intelligence methods,such as deep learning,can extract complex features from large amounts of data and adaptively process the data,providing new solutions for diverse prediction scenarios.The application of the Savitzky-Golay digital filter to the training set of the ultra-short-term wind power prediction model optimized the data and allowed the deep learning model to learn accurate time series patterns,thereby improving its generalization performance.The root-mean-square error for single-step prediction was reduced by 3.8% compared to the baseline model,and the errors for each stage of multi-step prediction were also smaller,resulting in improved accuracy and timeliness of the time series prediction model.The lead time of time series models for prediction was relatively shor t,but combining a power conversion model with NWP data c ould compensate for this limitation.Due to the complex and varied atmospheric conditions and the terrain where wind turbines were located,theoretical power models could easily produce deviations in predicting output power.However,power conversion models based on deep learning neural networks were more accurate.The RMSE between the theoretical power and measured power of the unit was 150.9117,and the SMAPE was 8.6850%.The RMSE between the power calculated by the deep learning model and the measured value was 106.7962,and the SMAPE was 7.7366%.When considering multiple feature variables,the RMSE of the deep learning model was 95.1779,and the SMAPE was 6.5611%,which further improved the accura cy of the prediction.Deep learning models had the advantage of efficient and accurate modeling of multi-dimensional non-linear systems,making them more adaptable to real-world scenarios.The mixture density network(MDN)based on neural networks and Gaus sian mixture models could quantitatively calculate the uncertainty of wind power.The Gaussian mixture model learned the distribution parameters automatically from the data during the neural network training process.The non-parametric nature of this architecture allowed it to better reflect the intrinsic structure of the data.The maximum expected value of the multivariate distribution output by the MDN model was used as the deterministic prediction result,with an n RMSE error of 0.0284 from the measured value,which was smaller than the error(n RMSE=0.0983)between the theoretical power of the unit and actual measured values.The MDN model could quantitatively describe the uncertainty of wind turbine output,providing more comprehensive information for scheduling planning and safety warning,and ensuring higher reliability for wind power operation and maintenance.When processing complex spatiotemporal sequences of meteorological parameters and energy data,models needed to be specially optimized.The Atten tionConv LSTM,designed for complex multidimensional spatiotemporal sequence data,employed the Conv LSTM architecture to extract and abstract local features while retaining sequence information,thereby improving the model’s expressive power.The model computed attention weights based on the output of the LSTM layer,allowing it to focus on key time steps and variable features,thereby improving the model’s accuracy.The Attention-Conv LSTM also used a residual network(Res Net)to ensure smooth gradient propagation during training.A comparison between hourly and daily prediction modes showed that the Attention-Conv LSTM performed better in handling complex multidimensional spatiotemporal sequence data.Taking into account the prediction of power generation an d demand,adjusting the output of thermal power units could reduce the proportion of traditional energy use.In two scheduling planning cases,the proportion of thermal power generation was reduced by 4.93% and 6.17% respectively,which meant more renewabl e energy could be effectively consumed and the proportion of wind and solar power abandonment was reduced by 14.7% and 16.23% respectively.Using the powerful data feature abstraction capabilities and flexible construction methods of deep learning neural networks,model improvements and functional extensions can be achieved,including multiple time scale calculations and uncertainty calculations,resulting in increased accuracy of source-load prediction.Based on source-load prediction,dispatch planning can adjust traditional energy output on the basis of prioritizing the consumption of renewable energy,thereby increasing the proportion of clean energy and protecting environmental resources.