Research On The Probabilistic Modeling And Prediction Method Of Source&Load Power Considering Temporal-spatial Correlation

In the next years,with the increasing share of renewable energy sources mainly of wind & solar power and new electric load with the uncertain space-time distributions,the bilateral randomness will be more noticeable in electricity supply & demand.It will present new challenges to the secure and economic operation of power systems.In this background,it has great academic value and practical significance to develop new probabilistic analysis and prediction models for accurately simulating and predicting the stochastic source & load power.Based on the different characteristics of WP/load in different weather,the study is carried out for time series probabilistic model of wind power(WP)and mid-term probabilistic prediction method of WP/load curve and WP curves from multi-wind farms by considering the randomness/fluctuation/temporal/spatial correlation of WP/load.It is supported by “Research on the basic theories of mid-long term probabilistic modeling and evaluation of transmission power system”,subsidized by the National Natural Science Foundation of China,No.51177178,and “Research on the integrated probabilistic planning method of TCTG and electrical collector system”,subsidized by the National Natural Science Foundation of China,No.51607014.Neglect of different WP characteristics in different weather by the traditional WP time series probabilistic models could result in inaccurately simulating the characteristics of WP and its change process with the weather.Aiming at this problem,a novel two-tier WP time series probabilistic model considering day-to-day weather transition and intraday WP fluctuations is proposed.The upper-tier model and the lower-tier model are established to characterize the day-to-day weather transition process and the intraday WP fluctuation processes in each typical weather state,respectively.In the upper-tier model,weather factors and conditions are classified into typical weather states in terms of the effects on daily average WP using a fuzzy clustering technique.Then a typical weather Markov chain model is established to characterize the day-to-day weather transition process.In the lower-tier model,in order to reduce accuracy dependence on power state number in the traditional Markov Chain Monte Carlo(MCMC)model,an improved MCMC model considering the probability distributions of the WP at the first time point of each day and WP fluctuations is developed to characterize the intraday WP fluctuation process.The proposed model is verified using the WP records and weather data at an actual wind farm.The results indicate that the proposed model can greatly improve the accuracy in capturing stochastic characteristics and time evolution characteristics of hourly and daily average WP without increasing power states.Concerning the mid-term weather forecasts,a mid-term probabilistic prediction method of source & load power curve combining factor analysis with quantile regression neural network(QRNN)probabilistic prediction method is proposed.It has lengthened the forecasting horizons of WP probabilistic prediction from current short term to middle term,and extended the single variable probabilistic prediction to multi-variables time series probabilistic prediction.Firstly,by means of the factor analysis model,the vector of 24 hourly WP/load seriesis decomposed into the loading matrix that reflects the correlation of 24 hourly power variables,common factors that contain the common features of the 24 original hourly power variables and specific components that contain the specific features of each original hourly power variable.Then,the independent common factors are used as the predictor variables and the QRNN models are built to predict the quantile series of each common factor separately.Concerning the difference of the related factors and temporal correlation of WP common factors between load common factors,the daily weather factors are addressed as the inputs of the WP common factor QRNN model,and the weather factors and weekday type of current day and the corresponding common factor of previous day are addressed as the inputs of the load common factor QRNN model.Also,in the training process of the load common factor QRNN model,the time distance weights of training samples are introduced into the error function estimating the model parameters to reflect different impacts of training samples with different distance to the predicted month.Finally,based on the prediction quantile series of common factors,the probability density functions of WP/load common factors are estimated by a kernel density estimation method.By simulating common factors and specific factors following the predicted distributions to recover WP/load prediction curves day by day,stochastic scenarios of the predicted WP/load in next several days are generated finally.The accuracy,adaptability and high efficiency of the proposed method have been verified by the predictions of actual wind farms with different capacities and system load.Based on the above mid-term probabilistic prediction method of WP power curve,a new mid-term WP probabilistic prediction method for multiple wind farms considering temporal-spatial correlation is proposed.It has extended the current mid-long term WP temporal-spatial correlation probabilistic modeling based on statistics theory to probabilistic prediction for a future day under specific weather condition.Based on the differences of linear temporal-spatial correlation in actual WP,3 typical factor models and their applications are presented to simplify temporal-spatial correlation of WP.Similarly,by concerning the mid-term weather forecasts and combining factor analysis with QRNN probabilistic prediction method,the mid-term WP curves of multiple wind farms are probabilistic predicted and randomly simulated.Finally,in order to solve the problem of huge amount of prediction scenarios,two technologies including the optimal discrete quantiles of common factor and K-medoids cluster are adopted to realize the scenario reduction and generate typical scenarios and the corresponding probabilities.The proposed methods and technologies are verified using the WP records and weather data at multiple actual wind farms.