Isolated micro-grid optimization method fusing high-proportion new energy and energy storage

Abstract

The invention discloses an isolated micro-grid optimization method fusing high-proportion new energy and energy storage, and belongs to the field of renewable energy and energy storage systems, and the method comprises the following three aspects: firstly, aiming at the optimal scheduling problem of an isolated micro-grid, formulating the optimal power scheduling problem of the isolated micro-grid into an MILP optimization model, and solving the optimal power scheduling problem of the isolated micro-grid; according to the model, detailed modeling of micro-grid components and alternating current power flow constraints is integrated, global optimality and feasibility are guaranteed, and the operation cost is minimized; and solving the optimization model by using historical data and an MILP solver to obtain a unit output data set of an optimal decision in the isolated microgrid. And secondly, training a dense residual neural network ResNetD on the obtained unit output data set, and deploying an SL strategy to learn and simulate an optimal ESS charging and discharging decision. And finally, applying the well trained ResNetD model, and providing an approximately optimal power dispatching decision according to real-time information.

Description

Technical Field

[0001] The present invention belongs to the field of renewable energy and energy storage systems, and in particular, relates to an isolated microgrid optimization method integrating a high proportion of new energy and energy storage. Background Art

[0002] Microgrids play an irreplaceable and important role and have great development potential in the flexible and efficient application of distributed power sources and new energy sources as well as intelligent control operations. Microgrids can be connected to external power grids for grid-connected operation, or disconnected from other power grids for isolated operation.

[0003] Isolated microgrids are a cost-effective means of providing electricity to remote or off-grid areas. However, due to the intermittency and uncertainty of renewable energy systems, isolated microgrids face significant challenges. In order to optimize the operation of isolated microgrids, proposing an optimal power dispatch strategy is essential to utilize renewable energy, reduce the impact of renewable energy interconnection, maintain a stable power supply, and reduce operating costs.

[0004] As a supplement to the centralized power grid, decentralized energy systems, isolated microgrids that integrate renewable energy sources (RES) and energy storage systems (ESS) have become key solutions beyond traditional grid connections. Optimal power dispatch is essential for the efficient operation, cost-effectiveness, and stability of isolated microgrids.

[0005] Current power dispatching methods such as stochastic optimization, robust optimization, interval optimization, incremental dynamic programming, and heuristic algorithms usually rely on accurate predictions of uncertain parameters. However, these predictions are biased, or the distribution information of uncertainty is inaccurate, and the resulting dispatching scheme may be suboptimal. This may lead to the need to adjust the already arranged dispatching plan or even re-dispatching it.

[0006] Reinforcement learning (RL) and deep reinforcement learning (DRL) algorithms are used in existing technologies to solve the power dispatch problem of isolated microgrids. Although these algorithms have strong optimization potential in theory, they usually need to learn through trial and error, which is not only time-consuming but also may encounter problems such as slow convergence, non-convergence or poor stability. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide an optimization method for an isolated microgrid integrating a high proportion of renewable energy and energy storage. The method proposes a new supervised learning (SL) strategy for optimizing the real-time power dispatching of isolated microgrids. The strategy significantly improves the processing speed, accuracy and stability by improving the existing residual neural network (ResNet) and optimization strategy.

[0008] To achieve the above purpose, the technical solution provided by the present invention is as follows: an isolated microgrid optimization method integrating a high proportion of new energy and energy storage, characterized in that it includes the following steps:

[0009] Step 1: With the goal of minimizing the operating cost, the optimal power dispatch problem of the isolated microgrid is formulated as a MILP optimization model. The MILP optimization is combined with the SL strategy. The MILP optimization model is solved by the MILP solver using historical data to obtain the unit output data set of the optimal decision in the isolated microgrid.

[0010] The power required by the isolated microgrid is provided by a diesel generator DEG and renewable energy generation, and the isolated microgrid operates under an incentive-based DR program; the renewable energy generation is solar photovoltaic PV generation and wind power generation WG generation, and an energy storage system ESS is used to store excess energy generated by renewable energy;

[0011] Step 2: Deploy the SL strategy to learn and simulate the optimal ESS charging / discharging decision by training a dense residual neural network ResNetD on the obtained unit output dataset;

[0012] Step 3: Use the trained dense residual neural network ResNetD to determine the optimal power dispatch decision based on real-time information.

[0013] According to a specific embodiment of the present invention, the power dispatch of an isolated microgrid is considered within a limited time range of 24 hours, T = 24, and a time interval of Δτ = 1, where Γ = {Δτ, 2Δτ, L, T} is set;

[0014] The power generated by the diesel generator DEG is represented by the operation limit in equation (1) and the ramp limit in equation (2);

[0015]

[0016] in, is the active power generated by the diesel generator in time interval t; p DEG and are the minimum and maximum dispatchable powers of the diesel generator respectively; is a binary variable representing the operating status of the diesel generator at time interval t, R DEG is the climbing limit of the diesel generator. is the active power generated by the diesel generator in the time interval t-1; is the active power generated by the diesel generator in the time interval t+1;

[0017] Available power from PV Modeled as a function of solar irradiance and ambient temperature, as follows:

[0018]

[0019] in and η PV are the rated active power and conversion efficiency of PV respectively; t is the solar irradiance at time interval t, θ t is the ambient temperature at time interval t;

[0020] The active power output of WG as a function of wind speed is defined as follows:

[0021]

[0022] in, is the active power generated by WG in time interval t; and η WG are the rated active power and conversion efficiency of WG respectively; γ t is the wind speed at time interval t; γ WG , γ WG,* and are the cut-in wind speed, rated wind speed and cut-out wind speed of WG respectively; α WG and β WG is the coefficient of WG;

[0023] The charging and discharging power of ESS are limited by the rated power, as shown in Equation (6) and Equation (7), respectively. In addition, the constraint given by Equation (8) forces the charging and discharging modes to be mutually exclusive.

[0024]

[0025]

[0026] in, and are the charging and discharging powers of ESS in time interval t, respectively; and are the rated charging and discharging powers of ESS in time interval t, respectively; and are binary variables representing the charging and discharging states of the ESS at time interval t, respectively;

[0027] The state of charge (SOC) of the ESS is modeled as a function of the system power exchange, as shown in Equation (9); in addition, the constraints defined by Equation (10) show that the SOC of the ESS is limited by the depth of discharge and the rated capacity;

[0028]

[0029] in, is the SOC of ESS at time interval t, is the SOC of ESS at time interval t-1; η ESS,ch and η ESS,dch are the charging and discharging efficiencies of ESS, respectively; ε ESS and are the minimum and maximum allowed capacities of the ESS, respectively;

[0030] Assume that the SOC of ESS reaches the maximum value in the first and last time intervals of the scheduling period, as shown below:

[0031]

[0032] The flexible load demand user model is as follows:

[0033]

[0034] in, and are the minimum and maximum load demands of the mth flexible load demand user that must be served in time interval t, is the load demand of the mth flexible load demand user after power cuts at time interval t, and M is the set of flexible load demand users;

[0035] The DistFlow equation is used to model the steady-state power flow in the power grid. The DistFlow equation is expressed as follows:

[0036]

[0037] Where N = {1,2,...,N B},E={1,2,...,N L},N B is the number of buses, N L is the number of lines; P ij|t and Q ij|t are the active power and reactive power from bus i to j in time interval t, P jk|t and Q jk|t are the active power and reactive power from bus j to bus k in time interval t; r ij and x ij are the resistance and reactance of line (i, j) respectively; and are the active and reactive power generation of bus j at time interval t, respectively; and are respectively the active and reactive power consumption of bus j at time interval t; set G = {DEG, PV, WG, ESS dch}, set C = {IL, FL1, L, FLm, L, FLM, ESS ch}, IL represents the rigid load demand user, FLm represents the mth flexible load demand user; v i|t is the square of the voltage amplitude on bus i at time interval t, l ij|t is the square of the complex current amplitude on line (i, j) at time interval t; in the power flow model, v i|t =|V i|t | 2 and l ij|t =|I ij|t | 2 , where V i|t and I ij|t are the complex voltage of bus i and the complex current on line (i, j) at time interval t; the voltage amplitude must satisfy equation (16), where v i and The lower and upper limits of the voltage amplitude are given respectively; in order to formulate the optimization problem as a convex problem, the linear quadratic equation in equation (17) is relaxed to the following second-order cone constraint:

[0038]

[0039] The operating cost of an isolated microgrid in time interval t is defined as follows:

[0040]

[0041] In formula (19), F t DEG represents the fuel and maintenance costs associated with DEG; the DEG cost can be defined as a quadratic function of the power generated by the DEG as follows:

[0042]

[0043] in, is the DEG cost coefficient;

[0044] F in formula (19) t OM represents the operation and maintenance costs of photovoltaic power generation, wind power generation and energy storage system; at the same time, the cost related to the energy storage system is expressed as a quadratic function of ESS using the formula:

[0045]

[0046] Among them, μ PV , μ WGand μ ESS are the operation and maintenance cost coefficients related to photovoltaic power generation, wind power generation and energy storage systems, is the charging / discharging power of ESS in time interval t;

[0047] When the microgrid operator needs to reduce the expected load demand, the flexible load demand users will receive compensation; the cost of applying the DR program is in the last term F in equation (19) t DR Consider and determine as follows:

[0048]

[0049] Among them, λ m is the price of the DR program for the mth flexible load demand user;

[0050] The optimal power dispatch problem is as follows:

[0051]

[0052] st:(1)-(22)

[0053] Among them, the decision variable matrix is ​​expressed as follows:

[0054]

[0055] in, Represent the reactive power of diesel generator, photovoltaic power generation and wind power generation respectively; Represents the reactive power consumed by the mth flexible load demand user.

[0056] According to a specific embodiment of the present invention, in the optimal power dispatch problem P1, due to the fluctuation of load demand and renewable energy generation, the SL strategy is adopted to partially transform the decision variables into action variables, and determine online based on the real-time information in the state variables. The state vector st describes the current state of the isolated microgrid: including the current time interval t, the output power of PV Output power of WG Rigid load requirements Flexible load demand and ESS in the previous period Therefore, the state vector is defined as follows:

[0057]

[0058] According to a specific embodiment of the present invention, the decision variables in equation (24) are defined in two steps:

[0059] In the first step, according to the observed state in equation (25), the SL strategy directly determines the charging / discharging power of the ESS at each time interval t, setting:

[0060]

[0061] in, They are the predicted values ​​of charging power, discharging power, charging state, discharging state, and SOC of ESS;

[0062] In the second step, a one-step optimization problem is performed for each time interval t, and the remaining decisions are determined as follows:

[0063]

[0064]

[0065] According to a specific embodiment of the present invention, the dispatching process of power dispatching in step 3 is as follows:

[0066] Phase 1: Generate unit output data sets by solving the MILP problem on historical data

[0067] In order to generate the unit output data set, the optimal power dispatch problem P1 is formulated as a deterministic MILP optimization model, which is solved by known input information; it is generated by collecting historical data or using scenarios; each scenario corresponds to a power dispatch cycle, and the historical data of a dispatch cycle is considered, with a period of T time intervals; the data includes rigid load demand users Flexible load demand user expectation value Flexible load demand user expectation value 2 Light intensity {υ Δτ ,υ 2Δτ ,...,υ T}, ambient temperature {θ Δτ ,θ 2Δτ ,...,θ T} and wind speed {γ Δτ ,γ 2Δτ ,...,γ T}; Use these data as input to the optimal power dispatch problem P1 and solve the problem as a day-ahead dispatch problem using a MILP solver;

[0068] By solving the optimal power dispatch problem P1, we obtain the input data and the optimal solution of each scenario to form T state-action pairs. A data set where the state vector s t Represents the observation information of each time interval t, the action vector a t Represents the state vector s tThe corresponding optimal action; the state-action pair for each time interval t is defined as follows:

[0069]

[0070] The state vector s in formula (29) t Rewritten from Equation (25), the load demand and renewable energy generation at time interval t are considered as one variable, namely, the net load, as shown in Equation (31); the action variable represents the charging / discharging power of the ESS, which is combined into a single variable using Equation (32); this iterative process is implemented for all considered scenarios, and the resulting data sets of state-action pairs are combined together; solving the optimal power dispatch problem P1 with N scenarios will produce a data set of NT state-action pairs,

[0071] Stage 2: Dense residual neural network ResNetD as a learning model

[0072] In the second stage, the goal is to learn a direct mapping π from the state of (29) to the corresponding optimal action of (30) * ; Therefore, the model is trained through SL using the state-action pair dataset; the input represents the state variables of the problem, and the output corresponds to their optimal behavior; by learning from these labeled data, the model is able to generalize and predict unseen inputs, thereby optimizing the solution in a data-driven manner; using the NT state-action pair dataset obtained by solving historical scenarios using the MILP solver in the first stage, the SL policy π maps a=π(s;θ), and the weight parameters θ are learned by training on the dataset D, which drives π towards the optimal policy π * near;

[0073] The network architecture of ResNetD architecture is: The network architecture of ResNetD architecture is: multiple blocks are connected in series to form the entire network structure, each block includes an input layer, a dense layer, a fully connected layer and an output layer arranged in sequence, wherein the dense layer includes at least two residual blocks, each residual block contains two convolutional layers and an activation function layer is arranged between the two convolutional layers, and each residual block is connected through a jump connection.

[0074] Phase 3: Real-time power dispatch

[0075] Real-time power dispatch at each time interval consists of two main stages; the first task is to find the optimal mapping π * The optimal ESS action is found under the condition that a well-trained ResNetD is used for real-time decision making; the proposed strategy collects real-time data and determines the state vector s at each time interval t. t ; Then, the trained ResNetDπ * With the state vector st As input, estimate the desired action The expected action in equation (30) To define the corresponding action of ESS; if the prediction has obvious deviation or violates the rated power limit or overcharge / discharge of ESS, the original output predicted by trained ResNetD is post-processed at each time interval t of the scheduling period to meet its related boundary constraints, as shown below:

[0076]

[0077]

[0078] Finally, by solving the one-step optimization problem P2, once the decision variables are fully determined, the system operator sends these values ​​as setpoint signals to the components of the isolated microgrid; this process is repeated at each time interval of the power dispatch cycle; therefore, the original optimization problem, the optimal power dispatch problem P1, is converted into a sequential decision problem, decomposed into multiple one-step optimization problems

[0079] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes an optimization method for isolated microgrids that integrates a high proportion of new energy and energy storage. Isolated microgrids that combine renewable energy and energy storage systems (ESS) have become a key solution for areas that are not connected to traditional power grids. Optimizing power dispatch is crucial to the efficient operation, cost-effectiveness and stability of isolated microgrids. Therefore, the present invention proposes a new supervised learning SL strategy for real-time optimal dispatch of isolated microgrids. The method is divided into three aspects: first, for the optimal dispatch problem of isolated microgrids, the optimal power dispatch problem of isolated microgrids is formulated as a MILP optimization model, which integrates detailed modeling of microgrid components and AC power flow constraints to ensure global optimality and feasibility to minimize operating costs; using historical data, the optimization model is solved using a MILP solver to obtain a unit output data set for optimal decisions in isolated microgrids. Secondly, by training a dense residual neural network (ResNetD) on the obtained unit output data set, the SL strategy is deployed to learn and simulate the optimal ESS charging and discharging decisions. Finally, a well-trained ResNetD model is applied to provide near-optimal power dispatch decisions based on real-time information. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a flow chart of an isolated microgrid optimization method integrating a high proportion of new energy and energy storage in an embodiment of the present invention;

[0081] Figure 2 Schematic diagram of an isolated microgrid in an embodiment of the present invention;

[0082] Figure 3 This is a schematic diagram of the control framework of the isolated microgrid system of the present invention;

[0083] Figure 4 This is the ResNetD architecture proposed in the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical scheme in the present invention is clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the present invention is not limited to the following embodiments, and the specific implementation method can be determined according to the technical scheme of the present invention and the actual situation. In order to avoid confusing the essence of the present invention, the known methods, processes and procedures are not described in detail.

[0085] The present invention provides an optimization method for an isolated microgrid integrating a high proportion of renewable energy and energy storage. The method proposes a new supervised learning SL strategy for real-time optimal power dispatch of an isolated microgrid integrated with RES and ESS to minimize the operating cost, such as Figure 1 shown.

[0086] An optimization method for isolated microgrids integrating a high proportion of renewable energy and energy storage is provided in the present invention based on direct learning strategies from expert demonstrations inspired by the behavioral cloning (BC) method. To achieve this goal, the power dispatch problem of isolated microgrids is initially defined as a mixed integer linear programming (MILP) optimization framework to minimize operating costs, where the DistFlow equation is applied to AC power flow modeling in power networks. Given historical data, the original power dispatch problem is solved as a deterministic day-ahead problem, and a MILP solver is used to obtain a dataset of unit output conditions for optimal decisions in isolated microgrids; subsequently, the proposed method learns and simulates the optimal ESS charging / discharging decisions by training a dense residual neural network (ResNetD) on the obtained dataset. The trained ResNetD is applied to provide near-optimal power dispatch decisions in real-time scenarios. The present invention proposes for the first time to develop a real-time power dispatch strategy for isolated microgrids based on the SL method.

[0087] Figure 2 Figure 1 is a schematic diagram of an isolated microgrid, which is not connected to the main grid. Therefore, the power required by the isolated microgrid is provided by a diesel generator (DEG) and renewable energy sources, including solar photovoltaic (PV) energy and wind power (WG) energy. In addition, an ESS is used to store excess energy generated by renewable energy. This stored energy can be used during periods of low renewable energy generation or peak demand. Figure 3The isolated microgrid system control framework diagram of the present invention is shown in Figure 1. The isolated microgrid operates under demand resources (DR) and considers the following two types of consumers:

[0088] Rigid load demand users: do not participate in the DR program, and user expectations must be fully met.

[0089] Flexible load demand users: Users whose expected demand can only be partially met will be compensated for every kWh of energy reduced.

[0090] In an isolated microgrid, the operations of various devices and local users are centrally managed. A communication network is deployed to connect signals between generators, consumers, ESS, and microgrid operators. Microgrid operators use energy management systems (EMS) to collect real-time data and implement real-time power dispatch strategies to minimize the operating costs of isolated microgrids. Dispatch signals are then transmitted to devices for optimal real-time operation.

[0091] The power dispatch of isolated microgrid is considered within a limited time range of 24 hours (T = 24) and a time interval of 1 hour (Δτ = 1), where Γ = {Δτ, 2Δτ, L, T} is set. The detailed mathematical expressions of each component of the isolated microgrid are elaborated in detail below.

[0092] The power generated by the diesel generator (DEG) is represented by the operating limit in equation (1) and the ramp limit in equation (2);

[0093]

[0094] in, is the active power generated by the diesel generator in time interval t; p DEG and are the minimum and maximum dispatchable powers of the diesel generator respectively; is a binary variable representing the operating status of the diesel generator at time interval t, R DEG is the climbing limit of the diesel generator. is the active power generated by the diesel generator in the time interval t-1; is the active power generated by the diesel generator in the time interval t+1.

[0095] Available power from PV Modeled as a function of solar irradiance and ambient temperature, as follows:

[0096]

[0097] in and η PVare the rated active power and conversion efficiency of PV respectively; t is the solar irradiance at time interval t, θ t is the ambient temperature at time interval t.

[0098] The active power output of PV must not exceed the peak power, as shown in equation (4).

[0099]

[0100] in, is the active power generated by the solar photovoltaic panel in time interval t.

[0101] The active power output of WG as a function of wind speed is defined as follows:

[0102]

[0103] in, is the active power generated by WG in time interval t; and η WG are the rated active power and conversion efficiency of WG respectively; γ t is the wind speed at time interval t; γ WG , γ WG,* and are the cut-in wind speed, rated wind speed and cut-out wind speed of WG respectively; α WG and β WG is the coefficient of WG.

[0104] Energy storage systems (ESS) play a vital role in maintaining a reliable and consistent power supply within isolated microgrids. The charging and discharging powers of energy storage systems are limited by the rated power, as shown in Equations (6) and (7), respectively. In addition, the constraints given by Equation (8) force the charging and discharging modes to be mutually exclusive.

[0105]

[0106] in, and are the charging and discharging powers of ESS in time interval t, respectively; and are the rated charging and discharging powers of ESS in time interval t, respectively; and are binary variables representing the charging and discharging states of the ESS at time interval t, respectively.

[0107] The state of charge (SOC) of the ESS is modeled as a function of the system power exchange, as shown in Equation (9). In addition, the constraints defined by Equation (10) show that the state of charge of the ESS is limited by the depth of discharge and the rated capacity.

[0108]

[0109] in, is the state of charge of the ESS at time interval t, is the charge state of the ESS at time interval t-1; η ESS,ch and η ESS,dch are the charging and discharging efficiencies of ESS, respectively; ε ESS and are the minimum and maximum allowed capacities of the ESS respectively.

[0110] Assume that the SOC of ESS reaches the maximum value in the first and last time intervals of the scheduling period, as shown below:

[0111]

[0112] DR programs are mechanisms that adjust electricity demand rather than supply and are divided into two main types: price-based DR programs and incentive-based DR programs. Price-based DR programs influence user electricity consumption by changing electricity prices over time. This strategy encourages users to reduce or shift electricity consumption to off-peak hours when electricity is cheaper and more abundant during peak electricity demand periods. At the same time, incentive-based DR programs provide price incentives to users to encourage them to voluntarily reduce electricity consumption during high demand periods or when grid reliability is at risk. The present invention deploys an incentive-based DR program in isolated microgrid modeling.

[0113] In this way, two types of consumers are considered in the isolated microgrid: rigid load demand users and flexible load demand users. For rigid load demand users, the active power and reactive power demand for each time interval t is given by and It indicates that the microgrid operator must fully meet these demands. At the same time, the flexible load demand users are willing to reduce a certain amount of expected demand according to the incentive-based DR program. Therefore, if the microgrid operator does not fully meet the user's expected demand, compensation will be paid to the user. However, the power that can be reduced must be within certain limits. Therefore, the flexible load demand users are modeled as follows:

[0114]

[0115] in, and are the minimum and maximum load demands of the mth flexible load demand user that must be served in time interval t, is the load demand of the mth flexible load demand user after power cut at time interval t, and M is the set of flexible load demand users. This embodiment considers one rigid load demand user (IL) and two flexible load demand users (FL1, FL2).

[0116] The present invention uses the DistFlow equation to model the steady-state power flow in the power grid. The proposed power flow model fully considers the active power and reactive power, branch current limitation and node voltage limitation, while basically maintaining the linearity of the equation. Therefore, the DistFlow equation is widely used in small-scale radial distribution networks, and its accuracy and reliability have been fully verified in many studies. The DistFlow equation is expressed as follows:

[0117]

[0118] Where N = {1,2,...,N B},E={1,2,...,N L},N B is the number of buses, N L is the number of lines; P ij|t and Q ij|t are the active power and reactive power from bus i to j in time interval t, P jk|t and Q jk|t are the active power and reactive power from bus j to bus k in time interval t; r ij and x ij are the resistance and reactance of line (i, j) respectively; and are the active and reactive power generation of bus j at time interval t, respectively; and are respectively the active and reactive power consumption of bus j at time interval t; set G = {DEG, PV, WG, ESS dch}, set C = {IL, FL1, FL2, ESS ch};v i|t is the square of the voltage amplitude on bus i at time interval t, l ij|t is the square of the complex current amplitude on line (i, j) at time interval t. In the power flow model, v i|t =|V i|t | 2 and l ij|t =|I ij|t | 2 , where V i|t and I ij|tare the complex voltage of bus i and the complex current on line (i, j) at time interval t. The voltage amplitude must satisfy equation (16), where v i and The lower and upper limits of the voltage amplitude are given respectively. In order to formulate the optimization problem as a convex problem, the linear quadratic equation in equation (17) is relaxed to the following second-order cone constraint:

[0119]

[0120] In the present invention, the operating cost of an isolated microgrid in time interval t is defined as follows:

[0121]

[0122] In formula (19), F t DEG represents the fuel and maintenance costs associated with DEG. The DEG cost can be defined as a quadratic function of the power produced by the DEG as follows:

[0123]

[0124] in, is the DEG cost coefficient.

[0125] F in formula (19) t OM represents the operation and maintenance costs of photovoltaic power generation, wind power generation and energy storage system. At the same time, the cost related to the energy storage system is expressed as a quadratic function of ESS using the formula:

[0126]

[0127] Among them, μ PV , μ WG and μ ESS are the operation and maintenance cost coefficients related to photovoltaic power generation, wind power generation and energy storage systems, is the charge / discharge power of the ESS in time interval t. In order to preserve the MILP formulation of the model, the quadratic terms in Eqs. (20) and (21) are linearized using piecewise linear approximation.

[0128] As mentioned above, when the microgrid operator needs to reduce the expected load demand, the flexible load demand users will receive compensation. The cost of applying the DR program is in the last term F in Eq. (19): t DR Consider and determine as follows:

[0129]

[0130] Among them, λ mis the price of the DR program for the mth flexible load demand user.

[0131] In the present invention, the optimal power dispatch problem involves determining the optimal dispatch of diesel generators, renewable energy, load shedding, and energy storage systems to meet the energy demand of the isolated microgrid at any given time, while considering the DR procedure and various system constraints. The main goal of the optimization problem is to minimize the operating cost of the isolated microgrid. To this end, PV and WG are prioritized for maximum utilization due to their operating costs and environmental benefits. When renewable energy generation is insufficient, the diesel generator is able to be controlled as a dispatchable source to fill the gap. The charge and discharge status of the ESS is optimized to store excess renewable energy and discharge it when there is insufficient generation or peak demand. In addition, an incentive-based DR procedure is implemented to adjust the load distribution according to the availability on the source side. Therefore, incentives are paid to encourage flexible users to shed load during peak hours or when renewable energy generation is low. The optimal power dispatch problem is as follows:

[0132]

[0133] st:(1)-(22)

[0134] Among them, the decision variable matrix is ​​expressed as follows:

[0135]

[0136] in, Represent the reactive power of diesel generator, photovoltaic power generation and wind power generation respectively; Represents the reactive power consumed by the mth flexible load demand user.

[0137] The optimization problem is subject to various constraints of the isolated microgrid. The power and ramp limits of the diesel generator are given in Equations (1) and (2), respectively. The maximum power generation of PV and WG is modeled in Equations (3), (4), and (5). The energy storage system charging / discharging constraints are defined in Equations (6), (7), and (8). The SOC of the energy storage system is modeled by Equation (9) and constrained by Equation (10), while the initial SOC and final SOC are fixed in Equation (11). Considering the DR procedure, the flexible load demand user modeling is defined by Equation (12). Finally, the AC power flow constraints are considered in Equations (13)-(18) for comprehensive modeling. All constraints must be satisfied for all time intervals.

[0138] The present invention proposes an isolated microgrid optimization method integrating high proportion of new energy and energy storage, which is mainly:

[0139] (1) By establishing a deterministic mixed integer linear programming (MILP) model and using historical data to obtain the unit output data set, a dense residual neural network (ResNetD) was trained to achieve fast and accurate power dispatching decisions.

[0140] (2) Combining MILP optimization with the SL strategy, the training efficiency is significantly improved and the training time is reduced by learning from the unit output data set calculated from historical data instead of starting from scratch.

[0141] (3) The present invention describes in detail the principles, specific implementation steps, key technical parameters, etc. of the combination of the MILP model and the SL strategy, including but not limited to the construction of the load forecasting model, the optimal scheduling strategy of the energy storage system, and the real-time control strategy of the microgrid.

[0142] (4) Through comparative experiments and data analysis, the superiority of the SL method in terms of performance, functionality, etc. Compared with the basic case, the SL method can reduce operating costs.

[0143] In the optimal power dispatch problem P1, due to the fluctuation of load demand and renewable energy generation, it is impossible to accurately predict the unit output and load in advance. To this end, a real-time power dispatch strategy based on supervised learning (SL) is proposed to minimize the operating cost of isolated microgrids. This strategy formulates the optimal power dispatch problem as a sequential decision problem with uncertainty in an unknown environment, partially converts the decision variables into action variables, and determines them online based on the real-time information in the state variables. The state vector s t Describe the current state of the isolated microgrid: including the current time interval t, PV output power Output power of WG Rigid load requirements Flexible load demand and ESS in the previous period Therefore, the state vector can be defined as follows:

[0144]

[0145] A simple approach is to use the optimized decision variables in (24) as action variables. However, when the number of actions is large over a long period of time, it is more challenging for the SL strategy to learn all actions. In addition, the learned actions may not satisfy all constraints. A more logical approach is to reduce the number of actions by selecting a characteristic variable of the power scheduling problem, i.e., the charging / discharging power of the ESS. Therefore, the decision variables in (24) are defined in two steps:

[0146] In the first step, according to the observed state in equation (25), the SL strategy directly determines the charging and discharging power of the ESS at each time interval t, assuming the following:

[0147]

[0148] in, They are the predicted values ​​of ESS’s charging power, discharging power, charging state, discharging state, and SOC.

[0149] In the second step, a one-step optimization problem is performed for each time interval t, and the remaining decisions are determined as follows:

[0150]

[0151] Its main advantage is that the action space processed by the SL strategy is greatly reduced. In addition, most constraints can be effectively handled by P2. After decomposing the problem into a sequential decision problem, the main purpose is to establish an SL strategy to make the best decision at each time interval t to minimize the operating cost of the isolated microgrid. Therefore, the first task is to generate a unit output data set consisting of state-action pairs, where the state vector represents the input information of the specific scenario of the problem, and the action vector represents the corresponding optimal solution obtained by explicitly solving the optimization problem. These state-action pairs serve as training data for model training. Therefore, the strategy proposes a ResNetD model that can capture the mapping relationship between state variables and corresponding optimal behaviors. After the model is fully trained, it can be used to predict or generate solutions that are close to the optimal solution, and can be used in the decision-making process of new scenarios. Since the time interval considered in the present invention is 1h (Δτ=1), the proposed real-time scheduling is regarded as a power scheduling strategy one hour ago.

[0152] The following is the scheduling process:

[0153] Phase 1: Generate unit output data sets by solving the MILP problem on historical data

[0154] In order to generate the unit output data set, the optimal power dispatch problem P1 is formulated as a deterministic MILP optimization model and accurately solved using precisely known input information. In practical applications, a large number of scenarios with accurate input information can be collected by collecting historical data or using scenario generation. Each scenario corresponds to a power dispatch cycle, and the historical data of a dispatch cycle is considered, with a period of T time intervals. The data includes rigid load demand users Flexible load demand user expectation value Flexible load demand user expectation value 2 Light intensity {υ Δτ ,υ 2Δτ ,...,υ T}, ambient temperature and wind speed {γ Δτ ,γ 2Δτ ,...,γ T}. These data are used as input to the optimal power scheduling problem P1, and the problem is solved as a day-ahead scheduling problem using the MILP solver. The optimal values ​​of the decision variables in equation (24) are obtained through the above process. However, the optimal solution obtained by the MILP solver represents an ideal solution (i.e., the best possible solution) because the optimal power scheduling problem P1 is formulated as a MILP problem and all information within a scheduling cycle is known in advance, which is impossible in actual operation.

[0155] By solving the optimal power scheduling problem P1, we can obtain the input data and the optimal solution of each scenario, and form T state-action pairs: A data set where the state vector s t Represents the observation information of each time interval t, the action vector a t Represents the state vector s t The corresponding optimal action. The state-action pair for each time interval t is defined as follows:

[0156]

[0157] The state vector s in formula (29) t Rewriting Equation (25), the load demand and renewable energy generation at time interval t are considered as one variable (i.e., net load), as shown in Equation (31). As mentioned above, the action variable represents the charging / discharging power of the ESS, which is combined into a single variable using Equation (32). This iterative process is implemented for all considered scenarios, and the resulting datasets of state-action pairs are combined together. Solving the optimal power scheduling problem P1 with N scenarios (each with T intervals) will produce a dataset of NT state-action pairs,

[0158] Phase 2: Dense Residual Neural Network (ResNetD) as a learning model

[0159] In the second stage, the goal is to learn a direct mapping π from the state of (29) to the corresponding optimal action of (30) *. Therefore, the model is trained by SL using the state-action pair dataset. The input represents the state variables of the problem, while the output corresponds to their optimal behavior (i.e., the optimal solution). By learning from these labeled data, the model is able to generalize and predict unseen inputs, thereby optimizing the solution in a data-driven manner. Using the NT state-action pair dataset obtained by solving historical scenarios using a MILP solver in the first stage, the SL policy π maps a=π(s;θ), and the weight parameters θ are learned by training on the dataset D. This process drives π towards the optimal policy π * Close. Technically, this problem is formulated as a regression problem with continuous inputs and outputs.

[0160] In traditional deep neural networks (DNNs), adding more layers poses challenges to the training of the network due to the problems of vanishing gradients and saturated training errors. However, in residual neural networks (ResNets), skip connections bypass some layers and perform identity mapping, where the output of one layer is added to the output of the layer from the previous steps. In this way, the gradient flow can be preserved throughout the network, allowing for deeper architectures without sacrificing performance. The present invention proposes a ResNetD as a learner model, which is inspired by the ResNet structure developed in previous studies. In the ResNetD architecture proposed in the present invention, as Figure 4 As shown, each block has at least two dense layers, which are merged by regular information flow, the output of the dense layer of the previous block, and direct connections from the input through the dense layer. The output of the last block is passed to the final output layer without activation. Figure 4 A ResNetD architecture with K=2 is proposed, where K is the number of convolutional layers W. i j The jth residual block of the i-th dense layer, X represents the input vector, Y represents the output vector, W-ReLU-W constitutes a dense layer, and the rectified linear unit (ReLU) is the activation function. By reusing the activation function of the previous layer until the adjacent layer learns its weights, these skipped connections are important in the "vanishing" and "exploding" gradient problems. Another benefit of skipping the layer is that it simplifies the network and speeds up the learning process because fewer layers are used in training, and multiple cascaded learning constitutes the ResNetD architecture.

[0161] Input layer: First, the sample scenery and load data are fed into the first convolutional layer W0 as input X.

[0162] Dense layer: The data then passes through a series of residual blocks. Each residual block contains two convolutional layers with an activation function layer between them to further extract features. Importantly, each residual block is connected via a skip connection.

[0163] Fully connected layer: After all residual blocks, the global average pooling layer reduces the spatial dimension of the data to generate a feature vector. This feature vector is then passed to the last convolutional layer W (also called a densely connected layer or linear layer) to perform classification or regression tasks.

[0164] Output layer: output predicted wind, light and load data.

[0165] This is typical for regression problems. The ResNetD proposed in the present invention uses residual connections (i.e., skip connections) to combine processed features and original input features at different stages, which enhances the gradient flow during training and has the potential to improve the learning of complex patterns in the data. ResNetD uses the rectified linear unit (ReLU) activation function and is trained using the Adam optimizer to minimize the mean squared error (MSE) loss function. The generalization ability of ResNetD is expected to provide a real-time power dispatching tool that can make near-optimal decisions in future scenarios. One block of the ResNetD architecture is to merge the regular information flow with the output of the dense layer of the previous block and directly connect the input (such as Figure 4 As shown in Figure 3, there are 2 skip neurons in the regular information flow). Since Haber loss is robust to outliers, Haber loss is adopted as the loss function. ReLU is used as the activation function of the proposed ResNetD.

[0166] Phase 3: Real-time power dispatch

[0167] The real-time power dispatch for each time interval consists of two main stages. The first task is to find the optimal mapping π * The proposed strategy collects real-time data and determines the state vector s at each time interval t. t . Then, the trained ResNetDπ * With the state vector s t As input to estimate the desired action The expected action in equation (30) To define the corresponding action of the ESS. If the prediction has a significant deviation or violates the rated power limit or overcharge / discharge of the ESS, the raw output predicted by the trained ResNetD is post-processed at each time interval t of the scheduling period to satisfy its related boundary constraints as follows:

[0168]

[0169]

[0170] Finally, by solving the one-step optimization problem P2, other variables such as the load demand of DEG and flexible customers can be obtained efficiently. Once the decision variables are fully determined, the system operator sends these values ​​as setpoint signals to the components of the isolated microgrid. This process is repeated at each time interval of the power dispatch cycle. Therefore, the original optimization problem (optimal power dispatch problem P1) is converted into a sequential decision problem and decomposed into multiple one-step optimization problems.

Claims

1. An optimization method for an isolated microgrid integrating a high proportion of new energy and energy storage, characterized in that: The following steps are involved: Step 1: With the goal of minimizing the operating cost, the optimal power dispatch problem of the isolated microgrid is formulated as a MILP optimization model. The MILP optimization is combined with the SL strategy. The MILP optimization model is solved by the MILP solver using historical data to obtain the unit output data set of the optimal decision in the isolated microgrid. The power required by the isolated microgrid is provided by a diesel generator DEG and renewable energy generation, and the isolated microgrid operates under an incentivized DR program; the renewable energy generation is solar photovoltaic PV generation and wind power generation WG generation, and an energy storage system ESS is used to store excess energy generated by renewable energy; Step 2: Deploy the SL strategy to learn and simulate the optimal ESS charging / discharging decision by training a dense residual neural network ResNetD on the obtained unit output dataset; Step 3: Use the trained dense residual neural network ResNetD to determine the optimal power dispatch decision based on real-time information.

2. The isolated microgrid optimization method integrating high proportion of new energy and energy storage according to claim 1 is characterized in that: The power dispatch of isolated microgrids is considered within a limited time range of 24 hours, T = 24, and a time interval of Δτ = 1, where Γ = {Δτ, 2Δτ, L, T} is set; The power generated by the diesel generator DEG is represented by the operation limit in equation (1) and the ramp limit in equation (2); in, is the active power generated by the diesel generator in time interval t; p DEG and are the minimum and maximum dispatchable powers of the diesel generator respectively; is a binary variable representing the operating status of the diesel generator at time interval t, R DEG is the climbing limit of the diesel generator. is the active power generated by the diesel generator in the time interval t-1; is the active power generated by the diesel generator in the time interval t+1; Available power from PV Modeled as a function of solar irradiance and ambient temperature, as follows: in and η PV are the rated active power and conversion efficiency of PV respectively; t is the solar irradiance at time interval t, θ t is the ambient temperature at time interval t; The active power output of WG as a function of wind speed is defined as follows: in, is the active power generated by WG in time interval t; and η WG are the rated active power and conversion efficiency of WG respectively; γ t is the wind speed at time interval t; γ WG , γ WG,* and are the cut-in wind speed, rated wind speed and cut-out wind speed of WG respectively; α WG and β WG is the coefficient of WG; The charging and discharging power of ESS are limited by the rated power, as shown in Equation (6) and Equation (7), respectively. In addition, the constraint given by Equation (8) forces the charging and discharging modes to be mutually exclusive. in, and are the charging and discharging powers of ESS in time interval t, respectively; and are the rated charging and discharging powers of ESS in time interval t, respectively; and are binary variables representing the charging and discharging states of the ESS at time interval t, respectively; The state of charge (SOC) of the ESS is modeled as a function of the system power exchange, as shown in Equation (9); in addition, the constraints defined by Equation (10) show that the SOC of the ESS is limited by the depth of discharge and the rated capacity; in, is the SOC of ESS at time interval t, is the SOC of ESS at time interval t-1; η ESS,ch and η ESS,dch are the charging and discharging efficiencies of ESS, respectively; ε ESS and are the minimum and maximum allowed capacities of the ESS, respectively; Assume that the SOC of ESS reaches the maximum value in the first and last time intervals of the scheduling period, as shown below: The flexible load demand user model is as follows: in, and are the minimum and maximum load demands of the mth flexible load demand user that must be served in time interval t, is the load demand of the mth flexible load demand user after power cuts at time interval t, and M is the set of flexible load demand users; The DistFlow equation is used to model the steady-state power flow in the power grid. The DistFlow equation is expressed as follows: Where N = {1,2,...,N B },E={1,2,...,N L },N B is the number of buses, N L is the number of lines; P ij|t and Q ij|t are the active power and reactive power from bus i to j in time interval t, P jk|t and Q jk|t are the active power and reactive power from bus j to bus k in time interval t; r ij and x ij are the resistance and reactance of line (i, j) respectively; and are the active and reactive power generation of bus j at time interval t, respectively; and are respectively the active and reactive power consumption of bus j at time interval t; set G = {DEG, PV, WG, ESS dch }, set C = {IL, FL1, L, FLm, L, FLM, ESS ch }, IL represents the rigid load demand user, FLm represents the mth flexible load demand user; v i|t is the square of the voltage amplitude on bus i at time interval t, l ij|t is the square of the complex current amplitude on line (i, j) at time interval t; in the power flow model, v i|t =|V i|t | 2 and l ij|t =|I ij|t | 2 , where V i|t and I ij|t are the complex voltage of bus i and the complex current on line (i, j) at time interval t; the voltage amplitude must satisfy equation (16), where v i and The lower and upper limits of the voltage amplitude are given respectively; in order to formulate the optimization problem as a convex problem, the linear quadratic equation in equation (17) is relaxed to the following second-order cone constraint: The operating cost of an isolated microgrid in time interval t is defined as follows: In formula (19), F t DEG represents the fuel and maintenance costs associated with DEG; the DEG cost can be defined as a quadratic function of the power produced by the DEG as follows: in, is the DEG cost coefficient; F in formula (19) t OM represents the operation and maintenance costs of photovoltaic power generation, wind power generation and energy storage system; at the same time, the cost related to the energy storage system is expressed as a quadratic function of ESS using the formula: Among them, μ PV , μ WG and μ ESS are the operation and maintenance cost coefficients related to photovoltaic power generation, wind power generation and energy storage systems, is the charging / discharging power of ESS in time interval t; When the microgrid operator needs to reduce the expected load demand, the flexible load demand users will receive compensation; the cost of applying the DR program is in the last term F in equation (19) t DR Consider and determine as follows: Among them, λ m is the price of the DR program for the mth flexible load demand user; The optimal power dispatch problem is as follows: st:(1)-(22) Among them, the decision variable matrix is ​​expressed as follows: in, Represent the reactive power of diesel generator, photovoltaic power generation and wind power generation respectively; Represents the reactive power consumed by the mth flexible load demand user.

3. The isolated microgrid optimization method integrating high proportion of new energy and energy storage according to claim 2 is characterized in that: In the optimal power dispatch problem P1, due to the fluctuations in load demand and renewable energy generation, the SL strategy is used to partially transform the decision variables into action variables, and determine them online based on the real-time information in the state variables. The state vector st describes the current state of the isolated microgrid: including the current time interval t, the output power of PV Output power of WG Rigid load requirements Flexible load demand and ESS in the previous period Therefore, the state vector is defined as follows:

4. The isolated microgrid optimization method integrating high proportion of new energy and energy storage according to claim 3 is characterized in that: The decision variables in equation (24) are defined in two steps: In the first step, according to the observed state in equation (25), the SL strategy directly determines the charging / discharging power of the ESS at each time interval t, setting: in, They are the predicted values ​​of charging power, discharging power, charging state, discharging state, and SOC of ESS; In the second step, a one-step optimization problem is performed for each time interval t, and the remaining decisions are determined as follows:

5. The isolated microgrid optimization method integrating high proportion of new energy and energy storage according to claim 4 is characterized in that: The dispatching process of power dispatch in step 3: Phase 1: Generate unit output data sets by solving the MILP problem on historical data In order to generate the unit output data set, the optimal power dispatch problem P1 is formulated as a deterministic MILP optimization model, which is solved by known input information; it is generated by collecting historical data or using scenarios; each scenario corresponds to a power dispatch cycle, and the historical data of a dispatch cycle is considered, with a period of T time intervals; the data includes rigid load demand users Flexible load demand user expectation value Flexible load demand user expectation value 2 Light intensity {υ Δτ ,υ 2Δτ ,...,υ T }, ambient temperature {θ Δτ ,θ 2Δτ ,...,θ T } and wind speed {γ Δτ ,γ 2Δτ ,...,γ T }; Use these data as input to the optimal power dispatch problem P1 and solve the problem as a day-ahead dispatch problem using the MILP solver; By solving the optimal power dispatch problem P1, we obtain the input data and the optimal solution of each scenario to form T state-action pairs. A data set where the state vector s t Represents the observation information of each time interval t, the action vector a t Represents the state vector s t The corresponding optimal action; the state-action pair for each time interval t is defined as follows: The state vector s in formula (29) t Rewritten from formula (25), the load demand and renewable energy generation in time interval t are regarded as one variable, namely, the net load, as shown in formula (31); The action variable represents the charging / discharging power of the ESS, which is combined into a single variable using equation (32); this iterative process is implemented for all considered scenarios, and the resulting data sets of state-action pairs are combined together; solving the optimal power dispatch problem P1 with N scenarios will produce a data set of NT state-action pairs, Stage 2: Dense residual neural network ResNetD as a learning model In the second stage, the goal is to learn a direct mapping π from the state of (29) to the corresponding optimal action of (30) * ; Therefore, the model is trained through SL using a dataset of state-action pairs; the input represents the state variables of the problem, and the output corresponds to their optimal behavior; By learning from these labeled data, the model is able to generalize and predict unseen inputs, thereby optimizing the solution in a data-driven manner; using the NT state-action pair dataset obtained by solving historical scenarios using the MILP solver in the first stage, the SL policy π maps a = π(s; θ), and learns the weight parameter θ by training on the dataset D. This process drives π towards the optimal policy π * near; The network architecture of ResNetD is as follows: multiple blocks are connected in series to form the entire network structure. Each block includes an input layer, a dense layer, a fully connected layer, and an output layer arranged in sequence. The dense layer includes at least two residual blocks. Each residual block contains two convolutional layers and an activation function layer is arranged between the two convolutional layers. Each residual block is connected by a jump connection. Phase 3: Real-time power dispatch Real-time power dispatch at each time interval consists of two main stages; the first task is to find the optimal mapping π * The optimal ESS action is found under the following conditions, where a well-trained ResNetD is used for real-time decision making; the proposed strategy collects real-time data and determines the state vector s at each time interval t t ; Then, the trained ResNetDπ * With the state vector s t As input, estimate the expected action The expected action in equation (30) To define the corresponding action of ESS; if the prediction has obvious deviation or violates the rated power limit or overcharge / discharge of ESS, the original output predicted by trained ResNetD is post-processed at each time interval t of the scheduling period to meet its related boundary constraints, as shown below: Finally, by solving the one-step optimization problem P2, once the decision variables are fully determined, the system operator sends these values ​​as setpoint signals to the components of the isolated microgrid; this process is repeated at each time interval of the power dispatch cycle; therefore, the original optimization problem, namely the optimal power dispatch problem P1, is converted into a sequential decision problem and decomposed into multiple one-step optimization problems.

Images