Optimal Intra-day Dispatch Of Smart Distribution System And Its Stochastic Dynamic Programming Policy

With the increasing penetration of renewable energy sources(RESs),energy storage systems(ESSs),micro grids(MGs)and demand side resources traditional distribution networks are undergoing a fundamental change to smart distribution networks.Due to the increasing stochasticity and deep coupling of different energy systems,traditional dispatch decision may be far from optimal or even infeasible.As a consequence,it is of great significance to investigate an online dispatch policy which adapts its decision to the revealed system state.Such an online dispatch policy can achieve more RES accommodation,and ensure the secure and economical operation of distribution networks.In essence,the optimal dispatch of smart distribution networks under multivariate uncertainties is typically a stochastic sequential decision problem with high dimensional and continuous-discreet state and decision space,and complex constraints.At present,the mostly-used intra-day dispatch suffers from dependence on accurate and comprehensive forecasts,computational complexity under multivariate uncertainties,and lose of optimality.Recently,the applications of intelligence approaches,e.g.,stochastic dynamic programming(SDP)and reinforcement learning,in complex decision-making problems have received great attentions.However,their difficulties in finding optimal policies under high dimensional complexity and infeasibility under complex constraints.This dissertation studies the optimal intra-day dispatch under multivariate uncertainties.The specific problems studied in this dissertation include convex and non-convex problems,centralized and decentralized dispatch framework,and electricity networks and multi-energy networks.This dissertation aims at developing a series of SDP methods which are computationally efficient,adaptive to stochasticity,and easily applicable.The major contributions are summarized as follows:1)By taking intra-day dispatch of smart distribution networks with high penetration of EVs as an example,this dissertation illustrates how distribution system operator(DSO)to utilize the flexibilities of distributed resources which are "large quantity and small capacity".Firstly,an equivalent model of large-scale EVs named EV clusters and hierarchical dispatch framework are proposed to tackle the difficulties and high dimensionality brought by uncertainties of EV owners' behaviors.Secondly,the SDP based optimal intra-day dispatch is given.By utilizing the value function concavity,a piece-wise linear value function approximation is employed and a temporal differential learning based policy search method is used.Simulation results verify the convergence,robustness,and computational efficiency of the proposed SDP policy.2)For the consideration of both economy and risk aversion,the optimal intra-day dispatch multi-energy smart distribution networks under dynamic risk aversion is investigated.Multivariate uncertainties are incorporated and multiple flexible resources are jointly optimized while satisfying networks constraints.Traditional SDP approaches which use expectation operator may suffer from the "small probability,high risk" events occurred in some scenarios with heavy tail distribution.By introducing risk-averse Markov decision process,an SDP based intra-day optimal dispatch with dynamic risk measure is proposed.Then,a data-driven based value function approximation is proposed to exploit knowledge from training samples.Thus,the almost optimal and computationally efficient policy is achieved.Simulation results illustrate that,compared with other popular methods,the proposed algorithm guarantees an economic and reliable dispatch for distribution networks.It not only facilitates economy but also reduces high quantile of total cost.It also helps in reducing RES curtailment and demand shedding.3)The optimal intra-day dispatch of multi-energy smart distribution networks with thermal storages is investigated.The original problem is a non-convex problem with multivariate uncertainties which can not be solved by the methods in 1)and 2).Firstly,the modelling of multi-energy smart distribution networks is explicitly formulated and the optimal day-ahead dispatch in deterministic scenarios is given.Secondly,an SDP based intra-day dispatch is proposed.Specifically,the value function monotonicity is proven and utilized to overcome "curse of dimensionality" of the original problem,then expert demonstrations are used to further accelerate offline learning process.Simulation results indicate that the proposed SDP achieves almost optimal decision,enjoys robustness.Moreover,the intra-day dispatch can be given in several seconds.4)Based on the problem in 3),a heuristics policy,which is easily applicable in practical and suitable for solving intra-day dispatch in complex power systems,is proposed.Firstly,an SDP policy using cost function approximation(CFA)is proposed.CFA converts difficult value function approximation problem to the optimization of the basis function parameters.Secondly,a novel imitation learning based policy search is proposed to produce optimal policy under a small amount of expert demonstrations.Finally,imitation learning problem is further formulated as a linear mixed-integer bilevel programs and a global solution is employed.Compared with SDP approach proposed in 3),the proposed method trades optimality and robustness for easy application and fast computation.5)To deal with heavy computation and communication burdens of traditional distributed optimization algorithms,the optimal distributed intra-day dispatch of smart distribution networks with multiple networked MGs using transactive energy control(TEC)is studied.Firstly,a TEC framework based on heterogenous decomposition is proposed and its optimality is rigorously proven.Secondly,a multi-agent SDP which is compatible with TEC framework is proposed and its optimality is also rigorously proven.Finally,a distributed offline self-learning algorithm is proposed to facilitate policy search without exchanging extra information among agents.After sufficient learning,multi-agent SDP can be executed in both non-iterative and iterative manners,which requires no iteration or only a few iterations among agents to obtain online solution.Besides,the policy provides great robustness to uncertainty and is very suitable for distributed decision-making in future smart grids.