Research On Ultra-short-term Prediction Of Photovoltaic Power Based On Improved Markov Model And Ground-based Cloud Images

In recent years,frequent extreme events threaten the reliability of the power system,while the randomness and volatility of renewable energy aggravate the difficulty of power grid regulation.Improving the prediction accuracy of wind and photovoltaic output is an important means to effectively guide the dispatch planning in the face of the current situation and reserve time for invoking adjustable resources,so as to improve the elasticity of power grid,reduce wind and light abandonment and promote energy consumption.Therefore,it is necessary to adjust with conventional power supply by improving the ultra-short-term perceptual regulation ability.At the same time,the rise of big data and artificial intelligence provides new ideas and tool support for prediction research.Under the new situation of high renewable energy penetration in power grid,the traditional forecasting technology will gradually fail to meet the demand.Considering the necessity and feasibility of improving the ability of ultra-short-term perceptual regulation,the research focuses on the power output prediction for the next 15 minutes to 4 hours.In this dissertation,the ultra-short-term prediction of photovoltaic power generation output under two scenarios is studied theoretically and verified by case studies.Several common methods of photovoltaic prediction have their own feature and applicable conditions.In the case of low time resolution and low accuracy of meteorological conditions,a Markov time series prediction method based on environmental feature clustering is proposed.The traditional Markov method is improved to optimize its state interval division method,and several sub-models are used to forecast different time periods.The influence of similar days and the number of state interval partition on the prediction result is analyzed.The case proves that the improved method can effectively improve the prediction effect.Since the cloud shielding effect on illumination radiation is the most direct factor of the photovoltaic power output fluctuation,the analysis of cloud image is very important for the ultra-short-term prediction.In another scenario,the prediction method based on the ground-based cloud image and recurrent neural network can be used for the station with meteorological data measurement equipment,especially the ability to collect cloud images.In the feature extraction of cloud image,the improved red-blue threshold ratio method is used to obtain the cloud coverage features.the main feature movement direction in the cloud image is calculated to obtain the brightness feature data of the key areas in the cloud image through SURF feature recognition and FLANN feature matching.The texture features of clouds in images are analyzed by Gabor filtering.The solar irradiation Angle of inclined plate was calculated and the nearby temperature data was obtained by querying.Using these characteristics and bidirectional LSTM network,the ultra-short-term prediction with a prediction step of 5 minutes is carried out.According to the prediction results,the improvement of cloud image features on the prediction results were analyzed,and the effects of the model parameters and structure changing on the prediction results were compared.The effectiveness of feature extraction and modeling was verified by the actual measured data of photovoltaic and cloud images obtained by the photovoltaic platform built on the roof.