Integrated Probabilistic Forecasting And Decision For Power System With Renewables

The significant intermittence and stochasticity of renewable energy generation have posed severe challenges to the secure and economic operation of power systems.Due to the chaotic nature of atmospheric systems,it is hard to predict the renewable energy generation without deviations,and the forecasting errors demonstrate extremely complex statistical characteristics.The energy balance between electricity supply and demand is thereby encountering considerable uncertainties.In the absence of reliable prognosis information,power system operators can hardly devise operational strategies with applicable feasibility and optimality.In this context,one of the key issues for renewable power systems is to accurately quantify the prediction uncertainties of renewable generation and establish the decision-making mechanism with high adaptability to prediction uncertainties.Relevant research area is meaningful to facilitate the large-scale integration and accommodation of renewable energy.Probabilistic forecasting is able to estimate possible values of renewable generation with the associated probability interpretations,which serves as the efficient tool for prediction uncertainty quantification and the indispensable input for power system decision-making models under uncertainties.Though with intensive attractions in both academic research and industrial practice,traditional probabilistic forecasting approaches commonly rely on prior parametric assumptions and can scarcely trade-off among multiple statistical performances.Moreover,the sequential forecasting and decision paradigm prevails in power systems,which neglects to optimize and evaluate the operational value of forecasting.To address these issues,this thesis investigates adaptive nonparametric interval forecasting and its applications in power system decision-making activities.The integrated probabilistic forecasting and decision methodologies are proposed in this thesis,which bridge the gap between probabilistic forecasting and decision from the perspectives of model formulation,information interaction,as well as performance optimization.The main contributions of this thesis are summarized as below:(1)An adaptive quantile regression approach is proposed for nonparametric interval forecasting.On the basis of quantile interpretations of prediction interval endpoints,an adaptive bilevel programming model is formulated,whose lower level problem generates a pair of calibrated quantile estimations by quantile regression and upper level problem tunes hyperparameters of quantile proportions to minimize the overall prediction interval width.The adaptive bilevel programming model is equivalently reformulated as a single level bilinear programming problem.An improved spatial branch-and-bound algorithm is developed to efficiently obtain the global optimum of the bilinear programming problem,which enhances the convex relaxation of bilinear constraint and convergence performance.The proposed adaptive quantile regression approach overcomes the conservativeness of interval width resulting from the symmetric restrictions of traditional central prediction intervals.By jointly optimizing the interval endpoints and the quantile proportion pair,the generated prediction intervals merit higher adaptivity to characterizing the asymmetric probability distributions of prediction uncertainty.(2)A chance constrained extreme learning machine approach is proposed for direct interval forecasting.The essential goal of probabilistic forecasting is to maximize the sharpness subject to well calibration.To this end,a chance constrained extreme learning machine model is formulated,which budgets the target coverage probability of prediction intervals by chance constraint and trains the extreme learning machine with the aim to minimize the expected interval width.The training process of the chance constrained extreme learning machine model is equivalent to a parameter searching task in zero-one loss minimization model subject to polyhedral feasible region.A difference of convex functions optimization based bisection search algorithm is developed to efficiently fulfill machine learning training by means of solving linear programming problems sequentially.The proposed chance constrained extreme learning machine approach is independent of fixed quantile restrictions on the prediction interval endpoints,and fully taps the latent potentialities for calibration and sharpness improvement.Various aspects of statistical performances for interval forecasting are balanced effectively.(3)A cost-oriented machine learning approach is proposed for interval forecasting and decision-making.Conventionally,probabilistic forecasting concentrates on enhancing the statistical performance but overlooks the operational value.The novel concept of cost-oriented prediction interval is proposed to minimize the decision cost brought by forecasting subject to well calibration.A cost-oriented machine learning framework is established with bilevel programming formulation.The lower level problems generate prediction interval endpoints and action plans,and upper level problem evaluates and optimizes the surrogate decision cost by adjusting quantile proportion pair.The bilevel programming model is tackled by the improved branch-and-bound algorithm,which simultaneously yields the sequences of cost-oriented prediction intervals and action plans.The proposed cost-oriented machine learning approach successfully bridges the information gap between probabilistic forecasting and decision making in renewable power systems,and remarkably uplifts the operational value of probabilistic forecasting.(4)An integrated probabilistic forecasting and decision approach is proposed for operating reserve quantification in power systems.According to the relationships between the bidirectional operating reserve requirements and prediction intervals of renewable generation,the interval forecasting and probabilistic reserve quantification processes are unified into a machine learning model.A loss function evaluating the reserve provision payment and deficit penalty is elaborated to realize cost-benefit trade-offs of reserve decision.Under the supervision of the loss function perceiving reserve cost,the machine learning model is established subject to the target coverage probability constraint of prediction intervals that guarantees the eligible confidence level of operating reserves.The integrated probabilistic forecasting and decision model for reserve quantification is formulated as a mixed integer linear programming problem.In order to reduce the computational complexity,a feasible region tightening strategy that shrinks the large constant coefficients and eliminates the redundant binary variables is developed to accelerate model training.The proposed integrated probabilistic forecasting and decision approach is capable of adjusting the prediction results to improve the decision performance,and achieves the collaborative optimization of probabilistic forecasting and decision performance.