An Active Distribution Network Integrated Optimization Scheduling Method and System Based on SVM-L2O

Abstract

This invention discloses an active distribution network integrated optimization scheduling method and system based on SVM-L2O, belonging to the field of distribution network optimization scheduling technology. The method includes: solving a multi-timescale integrated scheduling model based on acquired historical wind power output and load data to obtain unit start-up and shutdown strategies and continuous output decision variables, and constructing a training dataset; training a support vector machine for the start-up and shutdown strategies of each unit in each time period to obtain a surrogate model predicting the unit start-up and shutdown strategies; removing start-up and shutdown related constraints from the model based on the generated start-up and shutdown strategies to obtain a simplified model; constructing an L2O neural network, using Lagrange multipliers to incorporate the constraints of the simplified model as penalty terms into the loss function, training the L2O neural network to obtain a surrogate model predicting continuous output decision variables; acquiring wind power output and load data in real time, generating optimal scheduling decisions through the two surrogate models, and realizing integrated and efficient solution of multi-timescale distribution network scheduling.

Description

Technical Field

[0001] This invention relates to the field of distribution network optimization and scheduling technology, and in particular to an active distribution network integrated optimization and scheduling method and system based on SVM-L2O. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Renewable energy sources such as wind power and photovoltaics have significant advantages in zero emissions and sustainable development, and have become the core support for addressing the greenhouse effect and building a new type of power system oriented towards low-carbon transformation. They are playing an increasingly crucial role in the operation of modern active distribution networks.

[0004] However, under the existing day-day, intraday, and real-time hierarchical dispatching system, the high volatility and uncertainty of renewable energy output often lead to significant deviations in its power generation forecast curves across different time scales. This results in a weak correlation between day-day dispatching results and actual operating conditions at the intraday and real-time stages, thus reducing the guiding role of intraday dispatching in real-time grid operation. This imbalance in the connection between multiple time-scale dispatching stages not only weakens the practical reference value of dispatching plans at each stage but also adversely affects the economic efficiency and safety stability of active distribution network operation. Therefore, a new integrated dispatching model combining day-day, intraday, and real-time dispatching has been proposed. This model can simultaneously optimize all decision variables under the three time-scale dispatching, improving the system's adjustment flexibility and the economic efficiency and safety of distribution network operation.

[0005] However, the design of a multi-timescale integrated scheduling model will cause a significant surge in the number of decision variables and constraints, and the complexity of the model will increase exponentially. This will lead to excessively long solution time and heavy computational load, making it difficult to meet the rapid response requirements of high time resolution and high frequency rolling scheduling in the distribution network. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides an integrated optimization scheduling method and system for active distribution networks based on SVM-L2O. It establishes an integrated scheduling model for the distribution network that considers the impact of multiple time scales, including day-ahead, intraday, and real-time, and employs machine learning to construct a proxy model to support integrated and efficient solution of multi-time-scale scheduling of the distribution network. This solves the problems of incoordination and heavy computational burden in traditional multi-time-scale scheduling, effectively improving the timeliness and optimization capabilities of active distribution network scheduling and meeting the rapid response requirements of the distribution network.

[0007] In a first aspect, the present invention provides an active distribution network integrated optimization scheduling method based on SVM-L2O.

[0008] An active distribution network integrated optimization scheduling method based on SVM-L2O includes:

[0009] Historical wind power output and load data were acquired, a multi-time-scale integrated scheduling model of the active distribution network was built and solved, and the unit start-up and shutdown strategies and continuous output decision variables were obtained to construct a training dataset.

[0010] Based on the training dataset, support vector machines are trained for the start-up and shutdown strategies of each unit in each time period to obtain a proxy model for predicting the start-up and shutdown strategies of the units.

[0011] Based on the start-stop strategy generated by the proxy model, remove the start-stop related constraints in the model to obtain a simplified model;

[0012] An L2O neural network is constructed, and the constraints of the simplified model are incorporated into the loss function as a penalty term using Lagrange multipliers. The L2O neural network is trained using the training dataset to obtain a surrogate model that predicts the continuous output decision variables in the simplified model.

[0013] Real-time wind power output and load data are acquired, and optimal scheduling decisions for the active distribution network are generated through two surrogate models based on support vector machines and L2O neural networks.

[0014] A further technical solution is that the multi-time-scale integrated scheduling model aims to minimize the total operating cost of the distribution network. The total operating cost includes the cost of purchasing electricity from the upper-level grid, the operating cost of the generating units, the start-up and shutdown cost of the generating units, the standby cost of the generating units, the load shedding cost, and the cost of wind curtailment.

[0015] The constraints of the multi-timescale integrated scheduling model include power balance constraints, line flow constraints, line capacity constraints, unit start-stop logic variable constraints, minimum unit operation / outage time constraints, unit output constraints, unit ramp rate constraints, upstream grid power supply constraints, reserve constraints, node voltage constraints, wind curtailment and load shedding constraints, static var compensator constraints, unit start-stop state locking constraints, unit output locking constraints, and unit output maintenance time constraints.

[0016] A further technical solution involves using historical wind power output and load data, along with corresponding turbine start-up and shutdown strategy data, to train a support vector machine, including:

[0017] Using historical wind power output and load data as input and the start-stop status of the corresponding units as output, a support vector machine is trained for the start-stop strategy of each unit in each time period.

[0018] In the training process of support vector machines, the goal is to construct an optimal hyperplane that divides historical wind power output and load data and the corresponding start-up and shutdown states of the units into two categories: start-up and shutdown, and to maximize the distance between the hyperplane and the nearest point. The hyperplane parameters are determined by solving a linear optimization problem, and a mapping function is constructed based on the hyperplane parameters to learn the mapping relationship from wind power output and load characteristics to the start-up and shutdown states of the units.

[0019] A further technical solution is that the start-stop related constraints include unit start-stop logic variable constraints, unit minimum operating / shutdown time constraints, and unit start-stop state locking constraints.

[0020] A further technical solution is that the training process of the L2O neural network is as follows:

[0021] Historical wind power output and load data are acquired and combined with the unit start-up and shutdown strategies generated by the support vector machine proxy model to form the input dataset. At the same time, the continuous output decision variable data obtained from the multi-time-scale integrated scheduling model are extracted as the output dataset for supervision. Together, they constitute the training dataset of the L2O neural network.

[0022] Build an L2O neural network and train it using a training dataset to enable the network to learn the mapping from input to continuous output decision variables;

[0023] Specifically, Lagrange multipliers are used to transform the constraints of the simplified model into penalty terms that are incorporated into the loss function, thus constructing a total loss function based on the prediction error term and the constraint violation penalty term. Based on the total loss function, iterative training and parameter optimization of the L2O neural network are performed. By alternately updating the network parameters and Lagrange multipliers, the optimal parameters are obtained until the loss converges.

[0024] In a further technical solution, the prediction error term is the squared difference between the model prediction value and the supervised value, and the constraint violation penalty term is the weighted sum of the degree of violation of all equality and inequality constraints.

[0025] A further technical solution, optimizing the scheduling method, is as follows:

[0026] The real-time updated wind power output and load demand forecast data are input into the support vector machine surrogate model to obtain the start-up and shutdown strategies of all units; the start-up and shutdown strategies and real-time forecast data are input into the trained L2O neural network surrogate model to quickly output the continuous output decision variables.

[0027] The optimization scheduling method is optimized once every set time. The latest prediction data is used for optimization in each optimization cycle, and only the decision of the first scheduling period is executed in each optimization cycle. The scheduling plans of the remaining scheduling periods are used as references.

[0028] Secondly, this invention provides an active distribution network integrated optimization scheduling system based on SVM-L2O.

[0029] An active distribution network integrated optimization dispatching system based on SVM-L2O includes:

[0030] The model building module is used to acquire historical wind power output and load data, build and solve an integrated scheduling model of the active distribution network with multiple time scales, obtain unit start-up and shutdown strategies and continuous output decision variables, and build a training dataset.

[0031] The SVM training module is used to train support vector machines for the start-up and shutdown strategies of each unit in each time period based on the training dataset, so as to obtain a proxy model for predicting the start-up and shutdown strategies of the units.

[0032] The model simplification module is used to remove start-stop related constraints from the model based on the start-stop strategy generated by the proxy model, so as to obtain a simplified model.

[0033] The L2O neural network training module is used to construct an L2O neural network. It incorporates the constraints of the simplified model as a penalty term into the loss function using Lagrange multipliers, and trains the L2O neural network using the training dataset to obtain a surrogate model that predicts the continuous output decision variables in the simplified model.

[0034] The optimization scheduling module is used to acquire wind power output and load data in real time, and generates the optimal scheduling decision for the active distribution network through two surrogate models based on support vector machine and L2O neural network.

[0035] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the above-mentioned active distribution network integrated optimization scheduling method based on SVM-L2O.

[0036] Fourthly, the present invention also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the above-described active distribution network integrated optimization scheduling method based on SVM-L2O.

[0037] The above one or more technical solutions have the following beneficial effects:

[0038] This invention provides an active distribution network integrated optimization scheduling method and system based on SVM-L2O. It constructs an integrated distribution network scheduling model considering the influence of multiple time scales, including day-ahead, intraday, and real-time. Machine learning is used to build a surrogate model to support efficient integrated solutions for multi-time-scale distribution network scheduling. Specifically, support vector machines are used to quickly generate generator start-up and shutdown strategies based on real-time wind and solar power output and load demand data, avoiding the low efficiency caused by a large number of 0-1 variables in traditional methods. The Lagrange multiplier method is combined to incorporate constraints into the loss function of the L2O neural network, constructing an L2O neural network that enables rapid generation of continuous decisions such as generator output based on input load, wind power, and start-up / shutdown status. Finally, through the collaboration of the SVM and L2O surrogate models, rapid and efficient decision-making from real-time forecast data to a complete scheduling scheme is achieved, effectively solving the problems of slow variable generation and heavy computational burden in the integrated multi-time-scale model that combines day-ahead, intraday, and real-time scheduling. This invention transforms the multi-timescale integrated scheduling model from an optimization problem requiring online solution into a proxy model that can be trained offline and made rapid decisions online by constructing two proxy models, SVM and L2O. This can meet the rapid response requirements of high time resolution and high-frequency rolling scheduling in active distribution networks.

[0039] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is an overall flowchart of the active distribution network integrated optimization scheduling method based on SVM-L2O in Embodiment 1 of the present invention. Detailed Implementation

[0042] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] The overall concept proposed in this invention is as follows:

[0044] Considering the misalignment between day-ahead, intraday, and real-time scheduling, which reduces the reference value of scheduling plans at each stage and affects the economy and security of active distribution network operation, this invention designs an integrated distribution network scheduling model that simultaneously considers the influence of multiple time scales. This model integrates day-ahead, intraday, and real-time scheduling, and optimizes all decision variables under the original three time scale scheduling. It can fully tap the synergistic potential between power equipment, improve the system's regulation flexibility, effectively cope with the fluctuation problem of renewable energy, and improve the economy and security of distribution network operation.

[0045] Based on this, the high complexity of the aforementioned integrated model easily leads to problems such as excessively long solution time and heavy computational burden, failing to meet the rapid response requirements of high time resolution and high-frequency rolling scheduling in active distribution networks. Considering the excellent nonlinear mapping capabilities and data-driven characteristics of machine learning technology, a high-precision proxy model between input and output can be constructed through offline training, which can effectively approximate the physical correlation and optimization rules of complex scheduling models. Therefore, this invention uses machine learning to construct a proxy model to support the efficient solution of the integrated scheduling model of distribution networks at multiple time scales, improve the timeliness and optimization capabilities of active distribution network scheduling, and meet the rapid response requirements of distribution networks.

[0046] Example 1

[0047] This embodiment provides an active distribution network integrated optimization scheduling method based on SVM-L2O, such as... Figure 1 As shown, the specific steps include:

[0048] Step S1: Obtain historical wind power output and load data, build and solve an integrated scheduling model of the active distribution network with multiple time scales, obtain unit start-up and shutdown strategies and continuous output decision variables, and construct a training dataset.

[0049] Specifically, historical data on wind power output and load demand over multiple days are collected. Based on the constructed integrated multi-timescale scheduling model of the active distribution network (day-ahead, intraday, and real-time), the collected data is substituted to solve the scheduling problem for future periods, such as the next 24 hours, yielding a complete decision result including unit start-up and shutdown strategies and continuous output decision variables. Preferably, unit start-up and shutdown status data, as well as historical wind power output data and historical load demand data (i.e., historical load data) are extracted to form the training dataset for training the Support Vector Machine (SVM). The specific implementation process of step S1 above is as follows:

[0050] First, according to the set time granularity and scheduling cycle (Total number of scheduling periods) Collect historical predicted output data for each wind turbine in the power system and historical predicted demand data for each power load. The predicted power output data within a scheduling cycle is represented as follows: Electricity load Forecasted demand data within a scheduling period is represented as follows: and ,in, This represents active power demand forecast data. This represents reactive power demand forecast data, which in turn allows us to obtain historical datasets of all wind turbines and all power loads for each scheduling cycle. ,in and These represent the number of wind turbines and the number of electrical loads in the system, respectively. Select according to requirements. The data collected within a scheduling cycle can be used to construct the input dataset for training a support vector machine. .

[0051] Secondly, an integrated scheduling model for the active distribution network across multiple time scales (day-ahead, intraday, and real-time) is constructed. This model aims to minimize the total operating cost of the distribution network, taking into account costs of purchasing electricity from the upstream grid, unit operating costs, unit start-up and shutdown costs, unit standby costs, load shedding costs, and wind curtailment costs. Since the integrated scheduling model proposed in this embodiment has high-frequency rolling characteristics, a superscript is introduced. (n) Identifier n In this rolling process, to keep the model simple and easy to read, the index for decision variables, load included in the rolling process, and new energy forecast information is omitted unless otherwise specified. In this embodiment, the objective function of the multi-timescale integrated scheduling model can be expressed as:

[0052] (1.a)

[0053] In the above formula, This represents the total operating cost of the system within the scheduling period; These represent the costs of purchasing electricity from the upstream power grid, the operating costs of the generating units, the start-up and shutdown costs of the generating units, the standby costs, the load shedding costs, and the wind curtailment costs, respectively. These represent the index and total number of time periods, conventional generating units, power loads, and wind turbine units, respectively. This indicates the time granularity, which is set to 5 minutes in the multi-time-scale integrated scheduling model. For the upper-level power grid Electricity prices during the specified time period For the upper-level power grid Active power provided during the time period For the unit Operating cost coefficient, Indicates generator set In time period Active power within, Indicates the unit exist Binary variables representing the start / stop status within a time period. Indicates generator set Startup costs, Indicates that the unit is in the Binary variables that are started within a time period For the unit The alternative price, Indicates the unit In time period The backup provided inside Indicates the load shedding price. For electrical load In time period The reduction in active power load within the area. This represents the cost coefficient for wind curtailment. Indicates wind turbine In time period Power lost due to internal wind curtailment.

[0054] The constraints of the multi-timescale integrated scheduling model include:

[0055] (1) Power balance constraint, which represents the power flowing into the node Power and outflow nodes Since their power is equal, it can be expressed as:

[0056] (1.b)

[0057] (1.c)

[0058] In the above formula, These represent the index and total number of power lines and static var compensators, respectively. and They represent the lines respectively. The start and end nodes are relative to the nodes. The coupling coefficient matrix; These represent the upstream power grid, conventional generating units, wind turbines, and static var compensators relative to the nodes, respectively. The coupling coefficient matrix; and They represent Time-of-day routes Active power and reactive power; Indicates static var compensator exist The reactive power provided during the time period; and These represent wind turbine units. During the period Internally predicted active and reactive power; and Representing electrical load During the period Internally predicted active and reactive loads.

[0059] (2) Line power flow constraints, which take into account the coupling relationship between active power flow and reactive power flow in the distribution network. In this embodiment, the active power flow and reactive power flow constraint formulas are linearized as follows:

[0060] (1.d)

[0061] In the above formula, and They represent the lines respectively. The starting and ending nodes are nodes. ; and They represent the lines respectively. Conductivity and susceptance; and Representing nodes respectively exist Voltage and phase angle during the time period.

[0062] (3) Line capacity constraints can be expressed as:

[0063] (1.e)

[0064] In the above formula, Indicates the line The capacity.

[0065] (4) Unit start-up and shutdown logic variable constraints, used to describe the unit start-up and shutdown states and their changes, can be expressed as:

[0066] (1.f)

[0067] In the above formula, n This is the index representing the scrolling round; therefore, for each scrolling cycle, the index of the first time period is... t=n ; To identify the unit exist A binary variable indicating whether the unit has transitioned from a shutdown to an startup state. The transition from the shutdown state to the startup state is... ,otherwise ; To identify the unit exist A binary variable indicating whether the unit has transitioned from a startup to a shutdown state. The transition from the start state to the stop state is... ,otherwise ; Indicates the unit The initial running state.

[0068] (5) The minimum operating / downtime constraint of the unit can be expressed as:

[0069] (1.g)

[0070] (1.h)

[0071] In the above formula, and Similarly, it also represents an index for a time period; and They represent the generating units. The minimum operating time and minimum downtime (expressed in terms of time slots); Indicates the unit Initial cumulative operating time, i.e., unit The number of time periods that have been running continuously at the start of the scheduling cycle is a parameter that is continuously updated as optimizations are performed. The specific numerical value of the time represents the unit Number of time periods already running, The absolute value of the specific numerical value at that time represents the unit The number of time periods during which the unit has been shut down. Among them, equation (1.g) limits the minimum operating time of the unit, and equation (1.h) limits the minimum downtime of the unit.

[0072] (6) The unit output constraint can be expressed as:

[0073] (1.i)

[0074] In the above formula, and Generator sets The lower and upper limits of the active power output.

[0075] (7) The unit ramp rate constraint can be expressed as:

[0076] (1.j)

[0077] (1.k)

[0078] In the above formula, , They represent the generating units. The climb rate when in running state and the climb rate when in startup state. Indicates the unit At the initial power of the scheduling cycle, , They represent the generating units. The crawl rate when in operation and the crawl rate when shut down. Equations (1.j) and (1.k) describe the crawl rate limits of the unit.

[0079] (8) The power supply constraints of the upper-level power grid can be expressed as:

[0080] (1. l )

[0081] In the above formula, Indicates the upper-level power grid The upper limit of the active power that can be provided.

[0082] (9) The alternative constraint can be expressed as:

[0083] (1.m)

[0084] (1.n)

[0085] In the above formula, This represents the total reserve requirement for each time period. Equation (1.m) indicates that the reserve provided by the generating units is limited by the upper limit of output and the ramp rate, while Equation (1.n) indicates that the sum of the reserves provided by all generating units must meet the reserve requirement.

[0086] (10) Node voltage constraints can be expressed as:

[0087] (1.o)

[0088] In the above formula, and They are nodes The lower and upper limits of the voltage.

[0089] (11) Wind curtailment and load shedding constraints can be expressed as:

[0090] (1.p)

[0091] (1.q)

[0092] Equation (1.p) describes the upper and lower limits of wind curtailment by wind turbines, and Equation (1.q) describes the upper and lower limits of load shedding by power loads.

[0093] (12) The static var compensator constraint can be expressed as:

[0094] (1.r)

[0095] In the above formula, and These are the lower and upper limits of the reactive power output of the static var compensator, respectively.

[0096] (13) Unit start-up and shutdown status locking constraints

[0097] This indicates the cumulative number of rolling rounds during the rolling scheduling process, up to the set maximum number of locking rounds for the unit's start-up and shutdown status. Inside( The unit will be reset to 0 after reaching the maximum number of locking rounds. The start / stop state remains unchanged, that is, for satisfy:

[0098] (1.s)

[0099] In the above formula, Indicates the first During the second rolling optimization, the unit exist The start-up and shutdown status during a certain period. Equation (1.s) restricts the start-up and shutdown status of the unit to remain unchanged within a certain number of rolling cycles.

[0100] (14) Unit output locking constraint

[0101] This indicates the number of rolling cycles during which the unit output status is cumulatively locked, up to the set maximum number of locking cycles. Inside( The unit will be reset to 0 after reaching the maximum number of locking rounds. The output state remains unchanged, that is, for satisfy:

[0102] (1.t)

[0103] In the above formula, Indicates the first During the second rolling optimization, the unit exist The power output status during a certain period. Among them, equation (1.t) restricts the unit's power output status to remain unchanged within a certain number of rolling cycles.

[0104] (15) The unit output maintenance time constraint can be expressed as:

[0105] (1.u)

[0106] In the above formula, For the unit The output maintenance time. Wherein, equation (1.u) represents the unit's output remaining constant within a certain time range.

[0107] Furthermore, the historical datasets of all wind turbines and all power loads within the system will be used. Input the multi-timescale integrated scheduling model (1), and obtain the start-up and shutdown strategies of all conventional units for each time period through optimization. (superscript) This represents the optimal value obtained by optimizing the solver. Next, the selected... Input dataset within a scheduling cycle By inputting the multi-timescale integrated scheduling model (1) one by one according to the scheduling cycle, the output dataset of the corresponding training support vector machine can be obtained. This allows us to obtain the dataset needed to train the support vector machine. .

[0108] Step S2: Based on the training dataset, train a support vector machine for the start-up and shutdown strategy of each unit in each time period to obtain a surrogate model for predicting the start-up and shutdown strategy of the unit.

[0109] Specifically, after obtaining historical wind power output and load data and their corresponding generator start-up and shutdown decisions (as labels) in step S1, a training dataset is constructed. A support vector machine (SVM) is trained for the start-up and shutdown strategy of each conventional generator in each time period, learning the mapping relationship from wind and solar load characteristics to start-up and shutdown states. This enables the generator to quickly generate start-up and shutdown strategies based on real-time wind and solar load data, thus obtaining a surrogate model that can quickly output generator start-up and shutdown strategies based on real-time input prediction data. The specific implementation process of step S2 is as follows:

[0110] Since simultaneously deciding the start-up and shutdown status of multiple units at multiple times is a multi-classification problem, and its computational complexity increases rapidly with the number of units and time points, training a single support vector machine (SVM) is insufficient to achieve accurate and reliable unit start-up and shutdown decisions. Therefore, a simplified approach is adopted: a binary classifier is trained using an SVM for the start-up and shutdown strategy of each conventional unit at each time point to determine the unit's start-up and shutdown strategy.

[0111] For each conventional unit In each time period Each trained support vector machine can yield a different mapping function. To facilitate the training of the support vector machine, the training data... Rewritten Therefore, the set of all mapping functions can be obtained as follows: The corresponding sets of conventional generating units and time periods are respectively and .

[0112] In this embodiment, for each unit In each time period For each start-up and shutdown strategy, a linear support vector machine (SVM) can be developed to perform binary classification. Using historical wind power output and load data as input and the start-up and shutdown status of the corresponding turbine units as output, the SVM is trained using a training dataset. The training objective is to construct an optimal hyperplane that divides the training samples (historical wind power output and load data and their corresponding turbine start-up and shutdown statuses) into two classes: start-up / shutdown, and maximizes the distance between the hyperplane and the nearest point. To achieve this, the hyperplane parameters can be determined by solving a linear optimization problem.

[0113] (2.a)

[0114] (2.b)

[0115] In the above formula, and These are the coefficients that constitute the hyperplane. It is vectorized wind power and load data. It is a slack variable. Based on the multi-timescale integrated scheduling model, the scheduling problem for the next 24 hours is solved according to the rolling scheduling mode, obtaining unit start-up and shutdown data and continuous decision variable data. It is a vectorized representation of the unit start-up and shutdown data.

[0116] By solving problem (2), the optimal hyperplane coefficients can be obtained. and This allows us to construct mapping functions:

[0117] (3)

[0118] In the above formula, the function Indicates when hour , hour ; This represents the input feature vector; Indicates the output decision, i.e. , Indicates that the unit is in operation. This indicates that the generator unit has been shut down.

[0119] For each unit Each time period Each start-stop strategy solves problem (2) once, yielding the corresponding set of mapping functions. By simply inputting wind power output data and power load data, the start-up and shutdown strategies of all units in the system for all time periods can be obtained. In other words, through the above training process, a proxy model that generates start-up and shutdown strategies of all conventional units for all time periods in real time based on wind power output data and power load data can be obtained.

[0120] Step S3: Based on the start-stop strategy generated by the proxy model, remove the start-stop related constraints in the model to obtain a simplified model.

[0121] Specifically, the start-stop strategy proxy model trained in step S2 is used to generate start-stop strategies for all conventional units in each time period within the scheduling cycle. Then, combined with the complete start-stop strategy, the start-stop constraints in the original model are removed, resulting in a simplified multi-time-scale integrated scheduling model, which can be called the simplified model. The specific implementation process of step S3 is as follows:

[0122] The predicted wind power output data and load demand data within the scheduling cycle Input the set of mapping functions obtained from training step S2 That is, the support vector machine proxy model, which obtains the start-up and shutdown strategies of all conventional units in all time periods. To facilitate embedding into a multi-time-scale integrated scheduling model, Rewritten Then there is That is, to obtain each unit Each time period Start-stop decision By combining the unit start-up and shutdown logic variable constraints (1.f) in the multi-time-scale integrated scheduling model (1), the results for each unit can be obtained. In each time period Start / Stop state transition flag and Therefore, in the simplified model, the constraints of the unit start-up and shutdown logic variables (1.f) can no longer be considered. Furthermore, since the decision data is calculated through a complete multi-timescale integrated model, the resulting start-up and shutdown decisions... The minimum operating / shutdown time constraints (1.g) and (1.h) and the start-stop state locking constraint (1.s) of the unit can be naturally satisfied, so the influence of these constraints can be ignored in the simplified model.

[0123] By combining the start-stop decision agent model based on support vector machines, a simplified model of the multi-time-scale integrated scheduling model can be obtained, as follows:

[0124] (4.a)

[0125] (4.b)

[0126] (4.c)

[0127] (4.d)

[0128] (4.e)

[0129] (4.f)

[0130] (4.g)

[0131] (4.h)

[0132] (4.i)

[0133] (4.j)

[0134] (4.k)

[0135] (4.l)

[0136] (4.m)

[0137] (4.n)

[0138] (4.o)

[0139] (4.p)

[0140] (4.q)

[0141] Step S4: Construct an L2O neural network, use Lagrange multipliers to incorporate the constraints of the simplified model as penalty terms into the loss function, train the L2O neural network using the training dataset, and obtain a surrogate model for predicting the continuous output decision variables in the simplified model.

[0142] Specifically, an L2O neural network is constructed to predict the continuous decision variables in the simplified multi-timescale integrated scheduling model obtained in step S3. A physical residual loss function is constructed using the simplified integrated scheduling model after removing constraints related to unit start-up and shutdown. This involves incorporating the model's complex constraints as penalty terms into the loss function using Lagrange multipliers, forming the total loss function of the joint optimization prediction error and physical constraint loss. The learning optimization (L2O) neural network is then trained end-to-end using gradient descent, ensuring that the scheduling decision results of the output continuous decision variables can approach the optimal solution (i.e., minimize output cost) while strictly satisfying grid safety constraints. This achieves direct output of continuous decision variable scheduling instructions based on load, wind power forecasts, and predetermined start-up and shutdown states as input. The specific implementation process of step S4 is as follows:

[0143] First, for each scheduling cycle, input wind power output data and historical load demand dataset. Then, the agent model in step S3 generates start-up and shutdown decisions for the unit. Start-stop decisions throughout the entire scheduling cycle are Let the set of all unit start-up and shutdown data in each scheduling cycle be denoted as . By collecting the above data and combining it to form the input dataset, we can obtain the historical datasets of all wind turbines, power loads, and units for each scheduling cycle. Select Input dataset within a scheduling cycle As A training dataset is used. Simultaneously, continuous output decision variable data obtained from solving the multi-timescale integrated scheduling model are extracted and used as the output dataset for supervision, together forming the training dataset for the L2O neural network.

[0144] Secondly, construct an L2O neural network. (parameter is) The input to the L2O neural network is... The output of an L2O neural network is the predicted value of all continuous variables. ,Right now Based on each training sample Through the L2O model Calculate the predicted output .based on Calculate all constraint violations (power balance, ramp limit, voltage overrun, etc.) of the simplified integrated model.

[0145] Then, training is performed using Lagrange dual learning, with the training objective being to minimize the total loss while satisfying all constraints (4.b)-(4.q). This includes a prediction error term and a constraint violation penalty term, and can be expressed as:

[0146] (5.a)

[0147] The above prediction error term To measure the model's predictions With supervision value The gap between them. In step S1, a multi-timescale integrated scheduling model is constructed to solve the scheduling problem for the next 24 hours according to the rolling scheduling mode, obtaining unit start-up and shutdown data and continuous decision variable data, among which the continuous decision variable data is used as the supervision value for training the L2O model. , can be represented as:

[0148] (5.b)

[0149] Penalties for violating the above constraints To express the degree of violation of all equality and inequality constraints in the simplified model (4.b)-(4.q) of the multi-timescale integrated scheduling model, the weighted summation using Lagrange multipliers can be represented as:

[0150] (5.c)

[0151] in, This indicates an equality constraint (constraints (4.b)-(4.d), (4.p)-(4.q)); This indicates an inequality constraint (constraint (4.e) - (4.o)); For differentiable violated functions (use absolute value for equality, and inequality for inequality), use the function with ... ); It is a Lagrange multiplier, initialized to 0.

[0152] Next, an L2O neural network is trained using the training dataset, enabling the network to learn the mapping from input to continuous output decision variables. Specifically, iterative training and parameter optimization of the L2O neural network are performed based on the total loss function. This is achieved by alternately updating the network parameters and Lagrange multipliers until the loss converges, yielding the optimal parameters.

[0153] Specifically, the loss is calculated for the network parameters. The gradient is calculated, and the parameters of the L2O model are updated using gradient descent. :

[0154] (5.d)

[0155] in, is the learning rate. This step drives the network to adjust its output to find a solution that balances satisfying constraints and reducing costs.

[0156] At the end of each training cycle, use a fixed step size. Increase the Lagrange multipliers (penalty weights) of the violated constraints. Update equality constraints using equation (5.e) and inequality constraints using equation (5.f). This step gradually increases the penalty for violating constraints, driving the network output closer to the feasible region. The constraint is expressed as:

[0157] (5.e)

[0158] (5.f)

[0159] Repeat the above training process for multiple cycles, alternating between updating the network parameters. and the vehicle Continue this process until the loss converges. After training is complete, save the optimal parameters. and the vehicle The trained L2O neural network proxy model is obtained.

[0160] After training is complete, input the updated predicted wind power output data and load demand data. And the start / stop decision generated by the agent model in step S3. Let the input vector be... .Will Input to the trained L2O model This allows for the rapid output of continuous variables. The output solution automatically meets economic and constraint requirements without requiring online optimization problem solving. Through this framework, the multi-timescale integrated scheduling model can be transformed from an optimization problem requiring online solution into a neural network that can be trained offline and make rapid online decisions. Simultaneously, the Lagrange dual learning mechanism ensures the economic efficiency and feasibility of the solution.

[0161] Step S5: Acquire wind power output and load data in real time, and generate optimal scheduling decisions for the active distribution network through two proxy models based on support vector machine and L2O neural network.

[0162] Specifically, by collaborating with two agent models, SVM and L2O, a multi-timescale integrated scheduling model capable of rapid decision-making is obtained.

[0163] Through the above steps, two machine learning surrogate models are trained: a unit start-up and shutdown strategy generation model and a continuous decision variable generation model. By coordinating the operation of these two surrogate models (SVM and L2O), optimal scheduling decisions for integrated scheduling across multiple time scales can be quickly generated. This enables rapid and efficient decision-making from real-time forecast data to a complete scheduling plan, significantly improving the scheduling timeliness and optimization capabilities of the active distribution network. In this embodiment, the optimized scheduling method performs rolling optimization every 5 minutes, outputting all decision variables, including unit start-up and shutdown and various continuous variables, in each rolling optimization. Each optimization cycle uses the latest updated forecast information for optimization, and each optimization cycle contains multiple scheduling periods. After optimization, only the decision of the first scheduling period is binding and will be executed, while the scheduling plans for the remaining periods are only for reference.

[0164] Example 2

[0165] This embodiment provides an active distribution network integrated optimization scheduling system based on SVM-L2O, specifically including:

[0166] The model building module is used to acquire historical wind power output and load data, build and solve an integrated scheduling model of the active distribution network with multiple time scales, obtain unit start-up and shutdown strategies and continuous output decision variables, and build a training dataset.

[0167] The SVM training module is used to train support vector machines for the start-up and shutdown strategies of each unit in each time period based on the training dataset, so as to obtain a proxy model for predicting the start-up and shutdown strategies of the units.

[0168] The model simplification module is used to remove start-stop related constraints from the model based on the start-stop strategy generated by the proxy model, so as to obtain a simplified model.

[0169] The L2O neural network training module is used to construct an L2O neural network. It incorporates the constraints of the simplified model as a penalty term into the loss function using Lagrange multipliers, and trains the L2O neural network using the training dataset to obtain a surrogate model that predicts the continuous output decision variables in the simplified model.

[0170] The optimization scheduling module is used to acquire wind power output and load data in real time, and generates the optimal scheduling decision for the active distribution network through two surrogate models based on support vector machine and L2O neural network.

[0171] Example 3

[0172] This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0173] Example 4

[0174] This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0175] The steps involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0176] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0177] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. A method for integrated optimization scheduling of active distribution networks based on SVM-L2O, characterized in that, include: Historical wind power output and load data were acquired, a multi-time-scale integrated scheduling model of the active distribution network was built and solved, and the unit start-up and shutdown strategies and continuous output decision variables were obtained to construct a training dataset. The multi-timescale integrated scheduling model aims to minimize the total operating cost of the distribution network. The total operating cost includes the cost of purchasing electricity from the upper-level grid, the operating cost of generating units, the start-up and shutdown cost of generating units, the standby cost of generating units, the load shedding cost, and the cost of wind curtailment. The constraints of the multi-timescale integrated scheduling model include power balance constraints, line power flow constraints, line capacity constraints, unit start-stop logic variable constraints, minimum unit operation / outage time constraints, unit output constraints, unit ramp rate constraints, upstream grid power supply constraints, reserve constraints, node voltage constraints, wind curtailment and load shedding constraints, static var compensator constraints, unit start-stop state locking constraints, unit output locking constraints, and unit output maintenance time constraints. Based on the training dataset, support vector machines are trained for the start-up and shutdown strategies of each unit in each time period to obtain a support vector machine proxy model for predicting the start-up and shutdown strategies of the units. Based on the start-stop strategy generated by the proxy model, the start-stop related constraints in the model are removed to obtain a simplified model; the start-stop related constraints include unit start-stop logic variable constraints, unit minimum operating / outage time constraints, and unit start-stop state locking constraints. A learning optimization L2O neural network is constructed. The constraints of the simplified model are incorporated into the loss function as a penalty term using Lagrange multipliers. The L2O neural network is trained using the training dataset to obtain an L2O neural network surrogate model that predicts the continuous output decision variables in the simplified model. Real-time wind power output and load data are acquired, and optimal scheduling decisions for the active distribution network are generated using two surrogate models based on support vector machines and L2O neural networks. The optimized scheduling method is as follows: The real-time updated wind power output and load demand forecast data are input into the support vector machine surrogate model to obtain the start-up and shutdown strategies of all units; the start-up and shutdown strategies and real-time forecast data are input into the trained L2O neural network surrogate model to quickly output the continuous output decision variables. The optimization scheduling method is optimized once every set time. The latest prediction data is used for optimization in each optimization cycle, and only the decision of the first scheduling period is executed in each optimization cycle. The scheduling plans of the remaining scheduling periods are used as references.

2. The active distribution network integrated optimization scheduling method based on SVM-L2O as described in claim 1, characterized in that, Using historical wind power output and load data, along with corresponding turbine start-up and shutdown strategy data, a support vector machine is trained, including: Using historical wind power output and load data as input and the start-stop status of the corresponding units as output, a support vector machine is trained for the start-stop strategy of each unit in each time period. In the training process of support vector machines, the goal is to construct an optimal hyperplane that divides historical wind power output and load data and the corresponding start-up and shutdown states of the units into two categories: start-up and shutdown, and to maximize the distance between the hyperplane and the nearest point. The hyperplane parameters are determined by solving a linear optimization problem, and a mapping function is constructed based on the hyperplane parameters to learn the mapping relationship from wind power output and load characteristics to the start-up and shutdown states of the units.

3. The active distribution network integrated optimization scheduling method based on SVM-L2O as described in claim 1, characterized in that, The training process of the L2O neural network is as follows: Historical wind power output and load data are acquired and combined with the unit start-up and shutdown strategies generated by the support vector machine proxy model to form the input dataset. At the same time, the continuous output decision variable data obtained from the multi-time-scale integrated scheduling model are extracted as the output dataset for supervision. Together, they constitute the training dataset of the L2O neural network. Build an L2O neural network and train it using a training dataset to enable the network to learn the mapping from input to continuous output decision variables; Specifically, Lagrange multipliers are used to transform the constraints of the simplified model into penalty terms that are incorporated into the loss function, thus constructing a total loss function based on the prediction error term and the constraint violation penalty term. Based on the total loss function, iterative training and parameter optimization of the L2O neural network are performed. By alternately updating the network parameters and Lagrange multipliers, the optimal parameters are obtained until the loss converges.

4. The active distribution network integrated optimization scheduling method based on SVM-L2O as described in claim 3, characterized in that, The prediction error term is the squared difference between the model's predicted value and the supervised value, and the constraint violation penalty term is the weighted sum of the degree of violation of all equality and inequality constraints.

5. An active distribution network integrated optimization scheduling system based on SVM-L2O, characterized in that, include: The model building module is used to acquire historical wind power output and load data, build and solve an integrated scheduling model of the active distribution network with multiple time scales, obtain unit start-up and shutdown strategies and continuous output decision variables, and build a training dataset. The multi-timescale integrated scheduling model aims to minimize the total operating cost of the distribution network. The total operating cost includes the cost of purchasing electricity from the upper-level grid, the operating cost of generating units, the start-up and shutdown cost of generating units, the standby cost of generating units, the load shedding cost, and the cost of wind curtailment. The constraints of the multi-timescale integrated scheduling model include power balance constraints, line power flow constraints, line capacity constraints, unit start-stop logic variable constraints, minimum unit operation / outage time constraints, unit output constraints, unit ramp rate constraints, upstream grid power supply constraints, reserve constraints, node voltage constraints, wind curtailment and load shedding constraints, static var compensator constraints, unit start-stop state locking constraints, unit output locking constraints, and unit output maintenance time constraints. The SVM training module is used to train support vector machines for the start-up and shutdown strategies of each unit in each time period based on the training dataset, so as to obtain a support vector machine proxy model for predicting the start-up and shutdown strategies of the units. The model simplification module is used to remove start-stop related constraints from the model based on the start-stop strategy generated by the proxy model, so as to obtain a simplified model. The start-stop related constraints include unit start-stop logic variable constraints, unit minimum operating / outage time constraints, and unit start-stop state locking constraints. The L2O neural network training module is used to construct an L2O neural network. It incorporates the constraints of the simplified model as a penalty term into the loss function using Lagrange multipliers, and trains the L2O neural network using the training dataset to obtain an L2O neural network surrogate model that predicts the continuous output decision variables in the simplified model. The optimized scheduling module acquires wind power output and load data in real time. It generates optimal scheduling decisions for the active distribution network using two surrogate models based on support vector machines and L2O neural networks. The optimized scheduling method is as follows: The real-time updated wind power output and load demand forecast data are input into the support vector machine surrogate model to obtain the start-up and shutdown strategies of all units; the start-up and shutdown strategies and real-time forecast data are input into the trained L2O neural network surrogate model to quickly output the continuous output decision variables. The optimization scheduling method is optimized once every set time. The latest prediction data is used for optimization in each optimization cycle, and only the decision of the first scheduling period is executed in each optimization cycle. The scheduling plans of the remaining scheduling periods are used as references.

6. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the active distribution network integrated optimization scheduling method based on SVM-L2O as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The device stores executable instructions that, when executed by a processor, implement the active distribution network integrated optimization scheduling method based on SVM-L2O as described in any one of claims 1-4.

Images