A method for optimizing energy storage resources in a park building group

Abstract

The application discloses a kind of energy storage resource collaborative optimization methods for park building group, belong to energy storage resource optimization scheduling technical field.The method includes: from the real-time acquisition of multiple-source heterogeneous data of energy management platform of park building group, including the instantaneous load sequence of each building, photovoltaic power generation power sequence and the state of charge data of energy storage device;Multiple-source heterogeneous data is handled by time-space feature decoupling, generates the load fluctuation characteristic vector corresponding to each building and photovoltaic output confidence interval, and constructs the dynamic available capacity matrix of energy storage resource;Load fluctuation characteristic vector and dynamic available capacity matrix are input into dynamic game collaborative optimization model, and Nash equilibrium solution is solved by multiple rounds of iteration, generates the charge-discharge power distribution strategy of each energy storage device and the power interaction plan between buildings;According to charge-discharge power distribution strategy and power interaction plan, generate collaborative scheduling instruction set, issue to the energy storage controller of each building and execute.

Description

Technical Field

[0001] This invention relates to the field of energy storage resource optimization and scheduling technology, specifically to a collaborative optimization method for energy storage resources in a cluster of buildings in a park. Background Technology

[0002] Park buildings are typically equipped with distributed photovoltaic (PV) power generation systems and energy storage devices. By coordinating and optimizing the energy storage resources of each building, load fluctuations can be mitigated, PV absorption rates can be improved, and overall energy costs can be reduced. Existing energy storage resource optimization methods mostly adopt a centralized scheduling architecture, where a central controller collects load and power generation information from all buildings to solve for the globally optimal charging and discharging plan. This centralized approach ignores the differences in the preferences of each building as an independent stakeholder regarding its own energy storage usage rights. In actual operation, this can easily lead to conflicts of interest between buildings, making it difficult for optimization strategies to be accepted and stably implemented by each building. Existing methods also have shortcomings in handling the uncertainties of building load and PV power generation. Conventional practices usually only perform simple time-series smoothing or statistical analysis on load and PV power, without deeply exploring the frequency domain characteristics of load fluctuations or finely modeling the probability distribution of PV output. This results in coarse input information for the optimization model, making it difficult to adapt to real-time dynamic changes on both the source and load sides. When photovoltaic (PV) output fluctuates suddenly or load spikes occur, scheduling strategies based on coarse-grained characteristics are prone to problems such as improper allocation of energy storage capacity and unreasonable power interaction plans, affecting the overall economic efficiency and reliability of the park's operation. How to achieve real-time coordination of energy storage resources within a framework of balanced interests among distributed multi-stakeholder entities, and how to extract refined features from multi-source heterogeneous data that can characterize the dynamic changes in load and the uncertainty range of PV, are urgent problems to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a collaborative optimization method for energy storage resources in a cluster of buildings in a park. By treating each building as an independent player in a game, and introducing load frequency domain characteristics and photovoltaic probability prediction confidence intervals as model inputs, a globally executable energy storage charging and discharging strategy and inter-building power interaction plan are obtained while ensuring the balance of interests of each building.

[0004] To achieve the above objectives, this invention provides the following technical solution: This invention provides a collaborative optimization method for energy storage resources in a cluster of buildings within a park. By real-time collection of instantaneous load sequences, photovoltaic power generation sequences, and state-of-charge data of energy storage devices within the cluster, spatiotemporal feature decoupling processing is performed on the multi-source heterogeneous data to obtain the load fluctuation feature vectors of each building, the photovoltaic output confidence intervals, and the dynamic available capacity matrix of energy storage resources. This information is then input into a pre-constructed dynamic game-theoretic collaborative optimization model. After multiple rounds of iterative solving for the Nash equilibrium solution, a power allocation strategy for each energy storage device and a power interaction plan between buildings are generated. Finally, a collaborative scheduling instruction set is formed and issued for execution. This method can effectively integrate dispersed energy storage resources within a cluster of buildings, achieving reasonable power mutual assistance between buildings while considering load fluctuation mitigation and the uncertainty of photovoltaic absorption, thereby improving overall operational economy.

[0005] The spatiotemporal feature decoupling process for multi-source heterogeneous data specifically includes: discretizing the instantaneous load sequence according to a preset time granularity; performing a Fourier transform on the discrete load sequence to extract the amplitude components of each frequency band and sorting and combining them according to frequency to form a load fluctuation feature vector; inputting the photovoltaic power generation sequence into a probabilistic prediction model, generating multiple photovoltaic output scenarios using Monte Carlo sampling, and determining the upper and lower boundaries of the photovoltaic output confidence interval based on a given confidence level after statistical distribution fitting; obtaining the rated capacity and current state of charge of each energy storage device; subtracting the used capacity from the rated capacity and multiplying by the depth of discharge limitation factor to obtain the dischargeable capacity; subtracting the dischargeable capacity from the rated capacity to obtain the rechargeable capacity; and arranging the rechargeable and dischargeable capacities of all energy storage devices according to building number to construct a dynamic available capacity matrix. This processing method uniformly represents the frequency domain characteristics of the load, the uncertainty of photovoltaic output, and the real-time remaining capacity of energy storage, providing an accurate input benchmark for collaborative optimization.

[0006] In the dynamic game-theoretic collaborative optimization model, each building is considered a player in the game, the load fluctuation feature vector is considered private information, and the dynamic available capacity matrix is ​​considered global public information. First, the charging and discharging power allocation strategies of each player are initialized. Then, the gradient descent method is used to calculate the partial derivative of each player's objective function with respect to its own strategy, and its negative direction is used as the strategy adjustment direction. The strategy is updated along the adjustment direction at a preset step size. After boundary limiting processing, the total objective function of all players is recalculated until the difference between the total objective functions of two adjacent iterations converges to below a preset threshold, thus obtaining the Nash equilibrium solution. Based on this equilibrium solution, the difference in charging and discharging power among the players is used to determine the power interaction plan between buildings. This game-theoretic architecture allows each building to achieve system-level coordination while pursuing its own optimal goal, avoiding the communication and privacy burdens of centralized optimization.

[0007] Preferably, the objective function for each game participant is a linearly weighted sum of a load tracking deviation term, an energy storage loss cost term, and a power interaction cost term. The load tracking deviation term represents the difference between load fluctuation characteristics and charging / discharging power; the energy storage loss cost term reflects the ratio of charging / discharging power to available capacity; and the power interaction cost term is the absolute value of the difference between the participant's strategy and that of adjacent buildings. This objective function design allows the energy storage charging / discharging behavior to simultaneously achieve multiple effects, including local load peak shaving and valley filling, extending energy storage lifespan, and reducing interactive power losses.

[0008] As a technical solution of this invention, the power interaction plan is determined as follows: For adjacent building pairs with physical connections, the absolute value of the difference between the charging and discharging power of the two buildings in the Nash equilibrium solution is calculated as the power interaction demand. This demand is compared with the upper limit of the rated transmission power of the connection line, and the smaller value is taken as the power interaction scheduling value for the building pair. The power interaction plan is formed by summing the interaction directions and interaction power values ​​of all building pairs. This approach promotes energy sharing between buildings while strictly ensuring the operational safety of physical lines.

[0009] When generating the collaborative scheduling instruction set, the positive value portion of the charging and discharging power allocation strategy is used as the discharging instruction, and the negative value portion as the charging instruction. This is then superimposed with the corresponding interactive power value of the building in the power interaction plan to obtain a comprehensive power control instruction. This instruction is converted into a timed switching control signal sequence containing power setpoints and charging / discharging mode identifiers, encapsulated into a standard communication protocol message, and broadcast to the energy storage controllers of each building via the campus LAN for execution. This process achieves seamless conversion from optimization strategy to controller-executable instructions, ensuring the real-time performance and reliability of instruction issuance.

[0010] Preferably, the construction process of the dynamic game-theoretic collaborative optimization model includes: acquiring historical operation data of the park's building complex from multiple historical sampling periods, including historical load sequences, historical photovoltaic power generation sequences, and historical energy storage charging and discharging records; dividing the data into training and validation sets; training the probabilistic prediction model using the training set to obtain the trained photovoltaic power output prediction model; randomly selecting multiple initial strategy samples from the validation set, each sample corresponding to a set of randomly generated initial value matrices for charging and discharging power allocation, inputting them into the model to perform multiple rounds of iterative solutions, obtaining multiple converged Nash equilibrium solutions; calculating the objective function value corresponding to each equilibrium solution, and selecting the model parameters corresponding to the Nash equilibrium solution with the smallest objective function value as the final model parameters. This construction method utilizes historical data to optimize model parameters, enabling the model to have better convergence quality and global performance in actual operation.

[0011] Furthermore, when extracting initial policy samples from the validation set, the sample size is set to a positive integer greater than or equal to 50. N different initial value matrices for charge and discharge power allocation are generated by a uniform random number generator. The iterative solution is performed in parallel, and the convergence number and Nash equilibrium solution corresponding to each group are recorded. The solutions corresponding to the initial value matrices whose iteration convergence number exceeds the preset maximum iteration number are removed, and the remaining valid Nash equilibrium solutions are retained, thereby improving the robustness of the final model parameters.

[0012] The probabilistic prediction model preferably adopts a deep learning model based on a hybrid long short-term memory neural network and attention mechanism, which can accurately characterize the temporal features and abrupt changes of photovoltaic power output and provide a reliable confidence interval for photovoltaic power output.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By decoupling the spatiotemporal features of multi-source heterogeneous data, load fluctuation feature vectors and photovoltaic (PV) output confidence intervals are generated for each building. The instantaneous load sequence is discretized and then subjected to Fourier transform to extract the amplitude components of each frequency band. These components are then sorted and combined according to frequency to obtain the load fluctuation feature vector. This vector clearly separates the slowly varying base load component and the rapidly varying fluctuation component of the load from a frequency domain perspective, enabling the optimization model to formulate differentiated energy storage charging and discharging response strategies for different frequency band load components. The PV power generation sequence is input into a probabilistic prediction model constructed based on a hybrid long short-term memory neural network and attention mechanism. Multiple PV output scenarios are generated using Monte Carlo sampling and a statistical distribution is fitted. The upper and lower boundaries of the PV output confidence interval are determined by specifying a confidence level. This confidence interval reliably characterizes the uncertainty range of PV output, providing a probabilistic constraint on available PV power for subsequent game-theoretic optimization. Based on rated capacity, current state of charge, and depth of discharge constraint factors, a dynamic available capacity matrix is ​​constructed, consisting of dischargeable and rechargeable capacities arranged by building number. This matrix reflects the real adjustable capacity limit of each energy storage device at any given moment, ensuring that the optimization strategy is always solved within the physical safety range of energy storage. Each building is treated as a game player, with load fluctuation feature vectors serving as private information for each player and the dynamic available capacity matrix as global public information. A dynamic game-theoretic collaborative optimization model is constructed, and the Nash equilibrium solution is obtained. The objective function of each game player is a linearly weighted sum of load tracking deviation, energy storage loss cost, and power interaction cost. The gradient descent method is used to calculate the first-order partial derivative of the objective function with respect to its own charging and discharging strategy. The negative value of the partial derivative is used as the direction of strategy adjustment for multiple rounds of iterative updates until the overall objective function of all players tends to stabilize. This game-theoretic architecture enables each building to automatically reach an equilibrium state where no building can further reduce its own cost by unilaterally changing its strategy, while pursuing its own load tracking accuracy and minimizing energy storage loss and interaction costs. The obtained Nash equilibrium solution also provides the charging and discharging power allocation strategy for each energy storage device and the power interaction plan determined by the strategy difference between adjacent buildings and the transmission capacity of the physical connection line. Thus, under the condition of uncertainty on both the source and load sides and the coexistence of autonomous interests of multiple stakeholders, the efficient coordination of the overall energy storage resources of the park can be achieved. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of a collaborative optimization method for energy storage resources in a park's building complex. Figure 2 This is a flowchart of a dynamic game-theoretic collaborative optimization method; Figure 3 This is a flowchart of the construction process for a dynamic game-theoretic collaborative optimization model; Figure 4 This is a schematic diagram of the photovoltaic power output probability prediction model based on long short-term memory neural networks and attention mechanisms. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 This invention provides a collaborative optimization method for energy storage resources in a cluster of buildings within a park. The method comprises: real-time acquisition of multi-source heterogeneous data from the energy management platform of the cluster, including instantaneous load sequences, photovoltaic power generation sequences, and state-of-charge (SOC) data of each building; spatiotemporal decoupling processing of the multi-source heterogeneous data to generate load fluctuation feature vectors and photovoltaic output confidence intervals for each building, and constructing a dynamic available capacity matrix for energy storage resources; inputting the load fluctuation feature vectors and the dynamic available capacity matrix into a pre-constructed dynamic game-theoretic collaborative optimization model, and iteratively solving for Nash equilibrium solutions through multiple rounds to generate charging and discharging power allocation strategies for each energy storage device and power interaction plans between buildings; generating a collaborative scheduling instruction set based on the charging and discharging power allocation strategies and the power interaction plans, and distributing the collaborative scheduling instruction set to the energy storage controllers of each building for execution.

[0018] Example 1: In practice, spatiotemporal feature decoupling processing is performed on multi-source heterogeneous data to generate load fluctuation feature vectors and photovoltaic output confidence intervals for each building, and a dynamic available capacity matrix of energy storage resources is constructed.

[0019] The instantaneous load sequences of each building, collected in real time from the energy management platform of the park's building complex, are discretized according to a preset time granularity to generate discrete load sequences. The preset time granularity is set according to the data sampling period of the energy management platform, for example, 15 minutes. A Fourier transform is performed on the discrete load sequences to extract the amplitude components of each frequency band. The Fourier transform converts the discrete load sequences from the time domain to the frequency domain, outputting the complex amplitude of each frequency component. The modulus of the complex amplitude is taken to obtain the amplitude components of each frequency band. The amplitude components of each frequency band are sorted and combined according to frequency from high to low to generate a load fluctuation feature vector. The dimension of the load fluctuation feature vector is equal to the number of frequency bands involved in the sorting, arranged from low frequency to high frequency. The low-frequency amplitude components reflect the long-term trend of the load, while the high-frequency amplitude components reflect the short-term random fluctuations of the load.

[0020] In the process of generating the photovoltaic output confidence interval, the photovoltaic power generation sequence of each building is input into the probabilistic prediction model. The probabilistic prediction model is a deep learning model based on a hybrid long short-term memory neural network and an attention mechanism. The input of the probabilistic prediction model is the historical photovoltaic power generation sequence, and the output is the probability distribution parameters of the photovoltaic power generation in the future period. The specific structure of the probabilistic prediction model includes: a long short-term memory neural network layer, which performs time-series encoding on the input photovoltaic power generation sequence and extracts the hidden state sequence; an attention mechanism layer, which receives the hidden state sequence output by the long short-term memory neural network layer, calculates the attention score by using a trainable query vector and the hidden state sequence, and performs a weighted summation of the hidden state sequence to generate a context vector; and a fully connected output layer, which maps the context vector to probability distribution parameters, including the mean and variance of the predicted photovoltaic power generation value, which together define a normal distribution. The training process of the probabilistic prediction model includes: acquiring historical photovoltaic power generation sequences from multiple historical sampling periods as training samples, and dividing the training samples into training and validation sets; training the probabilistic prediction model using the training set, with negative log-likelihood loss as the loss function and Adam optimizer as the optimizer; evaluating the prediction interval coverage on the validation set, and stopping training when the interval coverage on the validation set deviates from the preset confidence level by less than a set threshold, thus obtaining the trained probabilistic prediction model.

[0021] When generating the photovoltaic output confidence interval, the photovoltaic power generation sequence before the current time is input into the trained probabilistic prediction model, and the probabilistic prediction model outputs the predicted mean of photovoltaic power generation for the next scheduling cycle. and prediction variance Using Monte Carlo sampling to obtain data from a normal distribution Generated in A photovoltaic power generation scenario The specific values ​​are determined based on computing resources. The value is 1000. (This is a partial translation of the original text, which is incomplete and requires further context.) Statistical distribution fitting was performed on several photovoltaic (PV) power output scenarios, and the upper and lower boundaries of the PV power output confidence interval were determined using the confidence level as a parameter. The confidence level was preset to 95%, so the lower boundary of the PV power output confidence interval was taken as... The 2.5th percentile of the scene value, with the upper bound being the 97.5th percentile.

[0022] When constructing a dynamic available capacity matrix for energy storage resources, the rated capacity, current state of charge (SOC), and depth of discharge (DOC) limit factor for each energy storage device are obtained. The rated capacity is obtained from the device's nameplate parameters; the SOC is read in real-time from the device's battery management system and expressed as a percentage; the DOC limit factor is the maximum allowable depth of discharge for the energy storage device, provided by the battery management system based on the battery's health status, with a value ranging from 0.7 to 0.9. The dischargeable capacity of each energy storage device is calculated as follows: subtract the used capacity corresponding to the current SOC from the rated capacity, and multiply the difference by the DOC limit factor. The used capacity equals the current SOC multiplied by the rated capacity. The dischargeable capacity of the energy storage system must consider both the current SOC and the DOC limit; the calculation formula is as follows: In the formula: Represents discharge capacity; Rated capacity (kWh); State of charge (dimensionless, range of values) ); This is the depth-of-discharge limitation factor (dimensionless, typically taken as 0.8~0.9). In the formula... This indicates the percentage of remaining usable capacity after deducting stored electricity, ensuring that the calculated value matches the actual physical capacity.

[0023] All energy storage devices are arranged in a matrix according to their building numbers to construct a dynamic available capacity matrix. Specifically, the rows of the matrix correspond to building numbers, which are determined based on the unified identification sequence of buildings within the park's building complex. The first column of the matrix stores the discharge capacity of the energy storage devices corresponding to each building, and the second column stores the recharge capacity. The dynamic available capacity matrix is ​​updated once before the start of each scheduling cycle to reflect the latest changes in state of charge.

[0024] Example 2: In specific implementation, please refer to Figure 2 The load fluctuation feature vector and dynamic available capacity matrix are input into a pre-constructed dynamic game collaborative optimization model. The Nash equilibrium solution is solved through multiple rounds of iteration to generate the charging and discharging power allocation strategy of each energy storage device and the power interaction plan between buildings.

[0025] Each building in the park's architectural complex is considered a player in a game, and the number of players is... , This equals the total number of buildings within the park's building complex. The load fluctuation feature vector of each building serves as private information for the corresponding game participant. This feature vector is composed of the frequency band amplitude components extracted from the discrete load sequences of each building after Fourier transform, sorted by frequency from highest to lowest. The dimension of the load fluctuation feature vector is denoted as... , This equals the number of frequency bands participating in the ranking. The dynamic available capacity matrix serves as global public information. Rows in the dynamic available capacity matrix correspond to building numbers, the first column stores the discharge capacity of the corresponding energy storage device, and the second column stores the rechargeable capacity. Initialize the initial charging and discharging power allocation strategy for each game participant. The initial charging and discharging power allocation strategy is a set of... A vector of n elements This equals the number of scheduling time windows. Each element in the vector represents the charging / discharging power value of the energy storage device within that time window; positive values ​​represent discharging, and negative values ​​represent charging. Each element of the initial charging / discharging power allocation strategy is set to zero.

[0026] Starting with the initial charging and discharging power allocation strategies of each game participant, the partial derivatives of each participant's objective function with respect to their own strategy are calculated using the gradient descent method. The direction of strategy adjustment is then determined based on these partial derivatives. First, the objective function for each game participant is constructed. The objective function is expressed as: in: Representing the participants in the game The objective function value; Representing the participants in the game The charging and discharging power allocation strategy vector, with a vector dimension of ; Representing the participants in the game The load fluctuation characteristic vector, with a vector dimension of ; Represents the Euclidean norm. For load tracking deviation; Represents the available capacity mapping function, when When one of the components in a dimension is greater than or equal to zero, the corresponding Take the game participants in the dynamic available capacity matrix The corresponding discharge capacity, when the component is less than zero, is... Take the game participants in the dynamic available capacity matrix The corresponding rechargeable capacity and energy storage loss cost are vector terms. Corresponding The sum of element-wise ratios; Representatives and Game Participants A set of adjacent buildings that are physically connected; Representing adjacent game participants The charging and discharging power allocation strategy vector; This is the power interaction cost item; The weighting coefficient representing the load tracking deviation term, with values ​​greater than 0. The weighting coefficient representing the energy storage loss cost item, with a value greater than 0. The weight coefficients representing the power interaction cost term have values ​​greater than 0. The three weight coefficients are determined through the parameter optimization process of the dynamic game collaborative optimization model, which is described in Example 4.

[0027] Computational game participants objective function Self-charging and discharging power allocation strategy The first-order partial derivative. For The first in The partial derivative consists of three terms: the first term is obtained by differentiating the load tracking deviation term, which is... The second term is obtained by differentiating the energy storage loss cost term. The time partial derivative is ,when The time partial derivative is ,in As a participant in the game Discharge capacity, As a participant in the game The rechargeable capacity; the third term is obtained by differentiating the power interaction cost term, which is... ,in Let be a sign function, taking the value 1 when the independent variable is positive, -1 when the independent variable is negative, and 0 when the independent variable is zero. The sum of the three partial derivatives is used as the game participants. The objective function of The first-order partial derivative is used as the direction of strategy adjustment, and the negative value of the first-order partial derivative is used as the direction of strategy adjustment.

[0028] The charging and discharging power allocation strategies of each game participant are updated along the strategy adjustment direction with a preset step size, and the overall objective function value of all game participants after the update is calculated. Preset step size. Based on the gradient descent convergence experiment, it was determined in advance. The value is set to 0.01, which balances the stability of the policy update magnitude and the convergence speed during the iteration process. The update method is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Current charge / discharge power allocation strategy The Each component plus the preset step size The product of the strategy adjustment direction and the updated charging / discharging power allocation strategy component values ​​is generated. After all components are updated, boundary limiting processing is applied to the updated charging / discharging power allocation strategy. The specific method of boundary limiting processing is as follows: check each updated component; if the component value is greater than zero and exceeds the limit of the game participants... Discharge capacity Then set the component to If the component value is less than zero and the absolute value exceeds that of the game participants... rechargeable capacity Then set the component to If the component value is within the allowable range, it remains unchanged. Boundary limiting processing ensures that the absolute value of the updated charge / discharge power allocation strategy does not exceed the upper and lower bounds of the corresponding element in the dynamic available capacity matrix.

[0029] The objective function for each game participant is recalculated based on the updated charging and discharging power allocation strategy, using the same calculation method as described above. The objective function values ​​of all game participants are then summed to generate the total objective function value. .

[0030] Determine whether the difference in the total objective function values ​​between two consecutive iterations is less than a preset convergence threshold. Preset convergence threshold Values This value is set based on the power control accuracy requirements within the scheduling cycle. The absolute value of the difference between the total objective function values ​​of two adjacent iterations is calculated; when this absolute value is less than... When convergence is reached, the charging and discharging power allocation strategies of each game participant obtained in the current iteration are determined as Nash equilibrium solutions. If the convergence condition is not met, the next iteration continues based on the updated charging and discharging power allocation strategies, repeating the process of calculating partial derivatives, determining adjustment directions, updating strategies, calculating the total objective function value, and determining convergence.

[0031] The power interaction plan is determined based on the charging and discharging power differences among the game participants in the Nash equilibrium solution. For each pair of adjacent buildings with physical connections, the charging and discharging power allocation strategies of the first and second buildings in the Nash equilibrium solution are obtained. The absolute value of the difference between the two strategies is calculated for each time window, and this absolute value is taken as the power interaction demand of that building for that time window. The rated transmission power limit of the physical connection line is obtained from the parameters of the park's electrical wiring diagram. The power interaction demand is compared with the rated transmission power limit, and the smaller value is taken as the power interaction scheduling value of that building for that time window. The interaction direction of the power interaction plan is determined according to the sign of the charging and discharging power difference: if the component of the charging and discharging power allocation strategy of the first building is greater than the corresponding component of the second building, the interaction direction is from the first building to the second building; if the component of the first building is less than the corresponding component of the second building, the interaction direction is from the second building to the first building. The power interaction scheduling values ​​and interaction directions of all building pairs in all time windows are summarized to generate a power interaction plan. The power interaction plan records the interaction direction identifier and interaction power magnitude of each building pair in each time window in matrix form.

[0032] Example 3: In practice, a collaborative scheduling instruction set is generated based on the charging and discharging power allocation strategy and power interaction plan, and then the collaborative scheduling instruction set is sent to the energy storage controllers of each building for execution.

[0033] The charging and discharging power allocation strategies of each game participant are extracted from the Nash equilibrium solution obtained by solving the dynamic game collaborative optimization model. The charging and discharging power allocation strategy is represented by a vector, where each component corresponds to a scheduling time window, and the value of the component represents the charging and discharging power command value of the energy storage device within the corresponding time window. The positive value part of the vector is used as the discharging command, and the negative value part is used as the charging command. That is, when the value of a component is greater than zero, a discharging command is generated, and the power value of the discharging command is the value of that component; when the value of a component is less than zero, a charging command is generated, and the power value of the charging command is the absolute value of that component.

[0034] The net interactive power value for each building in each time window is obtained from the power interaction plan. The power interaction plan records the interaction direction identifier and interaction power scheduling value for each building pair in each time window. For a given building, all building pairs containing that building are traversed, and the interaction direction identifier is read: if the interaction direction identifier indicates that interactive power flows from other buildings into this building, the corresponding interaction power scheduling value is added to the net interactive power value of this building; if the interaction direction identifier indicates that interactive power flows from this building to other buildings, the corresponding interaction power scheduling value is subtracted from the net interactive power value of this building. The net interactive power value is calculated separately for each time window; a positive value indicates that the building receives net power from the adjacent building group, and a negative value indicates that the building sends net power to the adjacent building group.

[0035] The charging / discharging commands and net interactive power values ​​of each building are superimposed to generate the comprehensive power control command for each building. The superposition calculation uses the following formula: in, Representative Buildings In the time window The comprehensive power control command is in kilowatts. A positive value indicates that the energy storage device is discharging to the outside, and a negative value indicates that the energy storage device is charging. Representative from architecture The charging and discharging command values ​​extracted from the charging and discharging power allocation strategy are in kilowatts. Positive values ​​correspond to discharging commands, and negative values ​​correspond to charging commands. Representative Buildings In the time window The net interactive power value, in kilowatts, with positive values ​​representing received power and negative values ​​representing output power.

[0036] The integrated power control command is converted into a sequence of timed switching control signals. For each time window, the timed switching control signal sequence includes a power setpoint and a charge / discharge mode identifier. The power setpoint is obtained from the integrated power control command. The absolute value, in kilowatts. Charge / discharge mode indication is based on... Assign values ​​to positive and negative values: If If the value is greater than zero, the charge / discharge mode indicator is set to discharge mode; if... If the value is less than zero, the charge / discharge mode indicator is set to charging mode; if... If the value is zero, the charging / discharging mode indicator is set to standby mode.

[0037] The timed switch control signal sequence is encapsulated into a standard communication protocol message. The standard communication protocol uses Modbus TCP, and the message data fields include the building number, timestamp, scheduling cycle start time, number of time windows, and the power setpoint value and charging / discharging mode identifier value for each time window. The number of time windows equals the total number of time periods divided within the scheduling cycle. The charging / discharging mode identifier is represented by a code: 1 for discharging mode, 2 for charging mode, and 0 for standby mode.

[0038] The encapsulated standard communication protocol messages are broadcast to the energy storage controllers in each building via the campus LAN. Each building's energy storage controller listens for these messages on the campus LAN, matching the building number carried in the message with its own preset building number, and only processes messages with matching building numbers. The energy storage controller parses the power setpoint and charge / discharge mode identifier for the time window sequence from the message and executes them sequentially according to the time window order: at the beginning of each time window, it sets the operating mode of the energy storage converter according to the charge / discharge mode identifier, sets the power command value to the corresponding power setpoint, and completes the charge / discharge power control for that time window.

[0039] Example 4: In specific implementation, please refer to Figure 3 The process of constructing a dynamic game-theoretic collaborative optimization model includes the following steps.

[0040] Historical operational data for the park's building complex is acquired across multiple historical sampling periods. This data is exported from the historical database of the park's building complex's energy management platform, with the total length of the historical sampling periods being no less than a complete annual operating cycle. The historical operational data includes historical load sequences, historical photovoltaic (PV) power generation sequences, and historical energy storage charge / discharge records. Historical load sequences are stored in time-series format, with each sampling point recording the active power value of each building, and the sampling interval consistent with the preset time granularity. Historical PV power generation sequences are also stored in time-series format, with each sampling point recording the output power value of each building's PV inverter. Historical energy storage charge / discharge records are stored in event log format, with each record including a timestamp, building number, charge / discharge power value, and charge / discharge mode.

[0041] Historical operational data was divided into training and validation sets based on time periods. The division method was as follows: using the start time of each historical sampling period as the starting point, the first 70% of the time period was used as the training set, and the last 30% as the validation set. The training and validation sets were continuous and non-overlapping on the time axis. The training set was used to train the probabilistic prediction model, and the validation set was used for parameter selection in the dynamic game collaborative optimization model.

[0042] Historical load sequences and historical photovoltaic power generation sequences from the training set are input into the probabilistic prediction model for training. The probabilistic prediction model is a deep learning model constructed using a hybrid long short-term memory neural network and an attention mechanism. The structure of the probabilistic prediction model includes: an input layer that receives a sequence of historical photovoltaic power generation sequences of length... The time window is used as input. The value is 96, corresponding to the number of sampling points divided into 15-minute granularities over 24 hours; a Long Short-Term Memory (LSTM) neural network layer containing 128 hidden units performs temporal encoding on the input sequence, outputting a hidden state vector at each time step, with a dimension of 128; an attention mechanism layer receives the sequence of hidden state vectors output by the LSM layer, using a trainable parameter vector. An inner product operation is performed with the hidden state vector at each time step. The softmax function is applied to the result to calculate the attention weights for each time step. The hidden state vectors are then summed according to the attention weights to generate a context vector with a dimension of 128. A fully connected output layer contains two output nodes, mapping the context vector to two output values. The first output value is used as the predicted mean of photovoltaic power generation. The second output value, after being processed by the softplus activation function, serves as the predicted variance of photovoltaic power generation. The softplus activation function ensures that the variance output is positive.

[0043] The probabilistic prediction model was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The loss function was negative log-likelihood loss, calculated as the negative of the log probability of the normal distribution determined by the mean and variance of the probabilistic prediction model's output at the actual photovoltaic power generation value. During training, the loss value on the validation set was calculated after each epoch. Training was terminated early if the validation set loss value did not decrease for 10 consecutive epochs, and the model parameters with the lowest validation set loss value were saved as the trained photovoltaic power output prediction model.

[0044] Multiple initial policy samples are randomly selected from the validation set. The number of initial policy samples is set to... indivual, The value is set to 50, which ensures sufficient parameter space search. A uniform random number generator is used to generate the parameters. A set of different initial value matrices for charge and discharge power allocation. A uniform random number generator in the interval... A uniformly distributed random number is generated. This random number is multiplied by the rated capacity of the corresponding building's energy storage device to serve as the initial value for charging and discharging power allocation. Each row of the initial charging and discharging power allocation matrix corresponds to the number of buildings, determined from the building list in the validation set data; the columns correspond to the number of scheduling time windows. , The value is 24, corresponding to a scheduling cycle with a granularity of 1 hour. The elements in the initial value matrix for each charge / discharge power allocation are represented as follows: , representing the The building is in The initial charge / discharge power values ​​for each time window.

[0045] Will The initial value matrices for charge / discharge power allocation are input into the dynamic game-theoretic collaborative optimization model, and multiple rounds of iterative solutions are performed in parallel. The process of these multiple rounds of iterative solutions is consistent with the iterative solution process used to generate the charge / discharge power allocation strategy and power interaction plan: each row of the initial value matrix is ​​used as the initial charge / discharge power allocation strategy for each game participant; starting from the initial charge / discharge power allocation strategy, the partial derivative of each game participant's objective function with respect to its own strategy is calculated using gradient descent to determine the direction of strategy adjustment; the strategy is updated along the adjustment direction with a preset step size and boundary limiting is applied; the updated total objective function value is calculated, and convergence is determined. (Preset step size) The value is 0.01, which is the preset convergence threshold. Values The maximum number of iterations is preset to 500.

[0046] Record the number of iterations and the final Nash equilibrium solution for each initial charge / discharge power allocation matrix. The number of iterations is the number of iterations performed by the dynamic game collaborative optimization model from the initial strategy to convergence. Remove Nash equilibrium solutions corresponding to initial charge / discharge power allocation matrices with more than 500 iterations, and retain the remaining solutions. A Nash equilibrium solution. It is a positive integer less than or equal to 50.

[0047] Calculate the remainder The objective function values ​​corresponding to each Nash equilibrium solution are as follows: based on the charging and discharging power allocation strategies of each game participant in a given Nash equilibrium solution, the sum of the objective function values ​​of all game participants is calculated by summing the total objective function value. The model parameters corresponding to the Nash equilibrium solution with the smallest total objective function value are selected as the final parameters of the dynamic game collaborative optimization model. The final parameters of the dynamic game collaborative optimization model include three parts: the weight coefficient of the load tracking deviation term. Weighting coefficient of energy storage loss cost item Weighting coefficient of power interaction cost term and the preset step size in gradient descent. During the screening process, , , The search scope is Each time the initial value matrix for charge / discharge power allocation is generated, the corresponding value is generated simultaneously. , , The three-parameter combination is obtained by random sampling from a log-uniform distribution. Specifically, the sampling method is as follows: within the interval... Generate uniformly distributed random numbers ,but In the interval Generate uniformly distributed random numbers ,but In the interval Generate uniformly distributed random numbers ,but The Nash equilibrium solution that minimizes the overall objective function value corresponds to... , , , The final parameters are incorporated into the dynamic game collaborative optimization model, thus completing the model construction.

[0048] Example 5: In specific implementation, please refer to Figure 4 The probabilistic prediction model is a deep learning model constructed based on a hybrid long short-term memory neural network and an attention mechanism. In the process of generating the photovoltaic output confidence interval, the model receives the photovoltaic power generation sequence as input, extracts temporal features through the long short-term memory neural network layer, and then weights the features at key time steps through the attention mechanism layer to output the probability distribution parameters of the photovoltaic power generation. This supports the generation of multiple photovoltaic output scenarios using the Monte Carlo sampling method, and then determines the upper and lower boundaries of the photovoltaic output confidence interval through statistical distribution fitting.

[0049] The probabilistic prediction model takes historical photovoltaic (PV) power generation sequences as input and outputs probability distribution parameters for future PV power generation. The length of the historical PV power generation sequence is... , The value is 96, corresponding to the photovoltaic power generation data of the past 24 hours sampled at 15-minute intervals. The input sequence is represented as a vector. ,in Represents the time step The photovoltaic power generation value is expressed in kilowatts.

[0050] The network structure of the probabilistic prediction model consists of three modules connected in series. The first module is a Long Short-Term Memory (LSTM) neural network layer, which contains 128 hidden units. The LTM neural network layer receives the input sequence. Processed sequentially by time step, at each time step Long Short-Term Memory (LSTM) neural network units based on the current input The hidden state vector of the previous time step and unit state vector Calculate the hidden state vector at the current time step. and unit state vector Hidden state vector The dimension is 128. The output of the Long Short-Term Memory (LSTM) neural network layer is a sequence of hidden state vectors for all time steps. .

[0051] The second module is the attention mechanism layer, which receives the sequence of hidden state vectors output by the long short-term memory neural network layer. The attention mechanism layer contains a trainable query vector. Query vector The dimension is 128, and the query vector is... The hidden state vectors are randomly generated during model initialization and updated during training via backpropagation. Calculate attention score The calculation method is as follows: in: Represents the query vector transpose, This represents the vector dot product operation. The scalar attention score.

[0052] Attention score over all time steps The attention weights are calculated using the softmax function. The softmax function is calculated as follows: ,in Represented by the natural constant An exponential function with base 0, and denominator 0. The sum of the exponential values ​​of the attention scores at each time step. Attention weights. satisfy The hidden state vector sequence By attention weight Perform a weighted summation to generate a context vector. Context vector The calculation method is as follows Context vector The dimension is 128.

[0053] The third module is a fully connected output layer, which receives the context vector. As input, the probability distribution parameters of the output photovoltaic power generation are used. The fully connected output layer contains a fully connected layer with two output nodes. The first output node directly outputs the predicted mean of the photovoltaic power generation. The second output node, after being processed by the softplus activation function, outputs the predicted variance of photovoltaic power generation. The expression for the softplus activation function is: ,in This is the original output value of the second output node of the fully connected layer. Represents the natural logarithm function. Represented by the natural constant An exponential function with a base of 0.5. The output of the softplus activation function is always positive, ensuring the prediction variance is... It is a positive value.

[0054] The training process for the probabilistic prediction model uses a training set partitioned from historical operational data. The historical photovoltaic power generation sequences in the training set are constructed as input-output sample pairs: for each time point, the previous time points are... The photovoltaic power generation values ​​of each sampling point are used as the input sequence, and the photovoltaic power generation value of the next sampling point after the current time point is used as the true label value. Training uses the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 100 training epochs. The loss function is negative log-likelihood loss. For a training sample, the negative log-likelihood loss is calculated as the mean of the predictions output by the probabilistic prediction model. and prediction variance The determined normal distribution in the actual photovoltaic power generation value The negative of the logarithmic probability at that point. Loss function The expression is: in: Represents pi (π) This represents the prediction variance output by the probabilistic prediction model. This represents the standard deviation of the forecast. The arithmetic square root, This represents the mean of the predictions output by the probabilistic prediction model. This represents the actual photovoltaic power generation value.

[0055] During training, the loss value on the validation set is calculated after each round of training. If the loss value on the validation set does not decrease for 10 consecutive rounds, the training is terminated early, and the model parameters with the lowest loss value on the validation set are saved as the probabilistic prediction model parameters after training.

[0056] When generating the confidence interval for photovoltaic output, the time before the current moment will be used as a reference. The photovoltaic power generation sequence of each sampling point is input into the trained probabilistic prediction model, and the probabilistic prediction model outputs the predicted mean. and prediction variance Monte Carlo sampling was used to sample samples from a normal distribution. Generated in A photovoltaic power generation scenario The value is 1000. (This is a partial translation of the original text, which is incomplete and requires further context.) The values ​​of each photovoltaic power output scenario are sorted from smallest to largest, and the value of the sorted scenario is taken. The first value is used as the lower boundary of the confidence interval for photovoltaic output, and the sorted value is taken as the first value. The values ​​are used as the upper bound, where Represents the confidence level. The value is 0.95. Represents the floor function. This represents the floor function. This statistical distribution fitting method directly estimates the quantile boundaries of the confidence interval from the generated scene data.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for collaborative optimization of energy storage resources for building complexes in industrial parks, characterized in that, The method includes: Multi-source heterogeneous data is collected in real time from the energy management platform of the park's building complex. The multi-source heterogeneous data includes the instantaneous load sequence of each building, the photovoltaic power generation sequence, and the state of charge data of the energy storage equipment. The multi-source heterogeneous data is subjected to spatiotemporal feature decoupling processing to generate load fluctuation feature vectors and photovoltaic output confidence intervals for each building, and a dynamic available capacity matrix of energy storage resources is constructed. The load fluctuation feature vector and the dynamic available capacity matrix are input into a pre-constructed dynamic game collaborative optimization model. The Nash equilibrium solution is solved through multiple rounds of iteration to generate the charging and discharging power allocation strategy of each energy storage device and the power interaction plan between buildings. Based on the charging and discharging power allocation strategy and the power interaction plan, a collaborative scheduling instruction set is generated and sent to the energy storage controllers of each building for execution.

2. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 1, is characterized in that, The process of decoupling the spatiotemporal features of the multi-source heterogeneous data to generate load fluctuation feature vectors and photovoltaic output confidence intervals for each building, and constructing a dynamic available capacity matrix for energy storage resources, specifically includes: The instantaneous load sequence is discretized according to a preset time granularity to generate a discrete load sequence. The discrete load sequence is then subjected to a Fourier transform to extract the amplitude components of each frequency band. The amplitude components are sorted and combined according to their frequency to generate the load fluctuation feature vector. The photovoltaic power generation sequence is input into the probabilistic prediction model, and multiple photovoltaic output scenarios are generated by the Monte Carlo sampling method. The multiple photovoltaic output scenarios are then subjected to statistical distribution fitting, and the upper and lower boundaries of the photovoltaic output confidence interval are determined using the confidence level as a parameter. The rated capacity and current state of charge of each energy storage device are obtained. The rated capacity is subtracted from the used capacity corresponding to the current state of charge, and then multiplied by the depth of discharge limitation factor to generate the dischargeable capacity of each energy storage device. The rated capacity is subtracted from the dischargeable capacity to generate the rechargeable capacity of each energy storage device. The dischargeable capacity and rechargeable capacity of all energy storage devices are arranged in a matrix according to the building number to construct the dynamic available capacity matrix.

3. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 2, is characterized in that... The process of inputting the load fluctuation feature vector and the dynamic available capacity matrix into a pre-constructed dynamic game-theoretic collaborative optimization model, and generating a charging and discharging power allocation strategy for each energy storage device and a power interaction plan between buildings through multiple rounds of iterative solutions to the Nash equilibrium, specifically includes: Each building is treated as a game participant, the load fluctuation feature vector is used as the private information of each game participant, and the dynamic available capacity matrix is ​​used as the global public information to initialize the initial charging and discharging power allocation strategy of each game participant. Starting with the initial charging and discharging power allocation strategy of each game participant, the partial derivative of the objective function of each game participant with respect to its own strategy is calculated by gradient descent method, and the direction of strategy adjustment is determined based on the partial derivative. The charging and discharging power allocation strategy of each game participant is updated along the strategy adjustment direction with a preset step size, and the total objective function value of all game participants after the update is calculated. Determine whether the difference between the total objective function values ​​of two adjacent iterations is less than a preset convergence threshold. If so, take the charging and discharging power allocation strategy of the current iteration as the Nash equilibrium solution, and determine the power interaction plan based on the charging and discharging power difference of each game participant in the Nash equilibrium solution.

4. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 3, is characterized in that... Starting with the initial charging and discharging power allocation strategies of each game participant, the method calculates the partial derivative of each participant's objective function with respect to its own strategy using gradient descent, and determines the direction of strategy adjustment based on the partial derivative. Specifically, this includes: Construct an objective function for each game participant. The objective function is composed of a linearly weighted load tracking deviation term, an energy storage loss cost term, and a power interaction cost term. The load tracking deviation term is the difference between the load fluctuation feature vector and the charging and discharging power allocation strategy. The energy storage loss cost term is the ratio of the charging and discharging power allocation strategy to the corresponding element in the dynamic available capacity matrix. The power interaction cost term is the absolute value of the difference between the charging and discharging power allocation strategy and the corresponding strategy of adjacent buildings. Calculate the first-order partial derivative of the objective function of each game participant with respect to their own charging and discharging power allocation strategy, and use the negative value of the first-order partial derivative as the direction of the strategy adjustment.

5. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 4, is characterized in that... The process of updating the charging and discharging power allocation strategy of each game participant along the strategy adjustment direction with a preset step size, and calculating the total objective function value of all game participants after the update, specifically includes: Add the product of the preset step size and the strategy adjustment direction to the current charging and discharging power allocation strategy of each game participant to generate an updated charging and discharging power allocation strategy, and perform boundary limiting processing on the updated charging and discharging power allocation strategy to ensure that its absolute value does not exceed the upper and lower bounds of the corresponding element in the dynamic available capacity matrix. Based on the updated charging and discharging power allocation strategy, the objective function of each game participant is recalculated, and the objective function values ​​of all game participants are summed to generate the total objective function value.

6. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 5, is characterized in that... The step of determining the power interaction plan based on the charging and discharging power differences among the game participants in the Nash equilibrium solution specifically includes: For each pair of adjacent buildings with physical connections, calculate the absolute value of the difference between the charging and discharging power allocation strategy of the first building and the charging and discharging power allocation strategy of the second building in the Nash equilibrium solution, and use the absolute value as the power interaction demand between the building pair. Obtain the rated transmission power limit of the physical connection line, compare the power interaction demand with the rated transmission power limit, and take the smaller value as the power interaction scheduling value for the building pair. The power interaction scheduling values ​​of all building pairs are summarized to generate the power interaction plan, which records the interaction direction and power magnitude of each building pair.

7. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 6, is characterized in that... The step of generating a collaborative scheduling instruction set based on the charging and discharging power allocation strategy and the power interaction plan, and then distributing the collaborative scheduling instruction set to the energy storage controllers of each building for execution, specifically includes: The positive value in the charging and discharging power allocation strategy is used as the discharging command, and the negative value is used as the charging command. The charging and discharging command of each building is superimposed with the interaction power value corresponding to that building in the power interaction plan to generate the comprehensive power control command for each building. The integrated power control command is converted into a timed switch control signal sequence, which includes the power setpoint and charge / discharge mode identifier for each time window. The timed switch control signal sequence is encapsulated into a standard communication protocol message and broadcast to the energy storage controllers of each building via the park's local area network.

8. The method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 7, is characterized in that... The method further includes constructing the dynamic game-theoretic collaborative optimization model, specifically including: The historical operation data of the park's building complex was obtained from multiple historical sampling periods. The historical operation data included historical load sequences, historical photovoltaic power generation sequences, and historical energy storage charging and discharging records. The historical operating data is divided into training set and validation set according to time period. The historical load sequence and historical photovoltaic power generation sequence in the training set are input into the probabilistic prediction model for training, and a trained photovoltaic output prediction model is generated. Multiple initial strategy samples are randomly selected from the verification set. Each initial strategy sample corresponds to a set of randomly generated initial values ​​for charging and discharging power allocation. The initial strategy samples are input into the dynamic game collaborative optimization model to perform multiple rounds of iterative solution to obtain multiple converged Nash equilibrium solutions. Calculate the objective function value corresponding to each Nash equilibrium solution, and select the model parameters corresponding to the Nash equilibrium solution with the smallest objective function value as the final parameters of the dynamic game collaborative optimization model.

9. A method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 8, is characterized in that, The process involves randomly selecting multiple initial strategy samples from the validation set, each initial strategy sample corresponding to a set of randomly generated initial values ​​for charging and discharging power allocation. These initial strategy samples are then input into the dynamic game-theoretic collaborative optimization model to perform multiple rounds of iterative solving, obtaining multiple converged Nash equilibrium solutions. Specifically, this includes: The number of initial strategy samples is set to N, where N is a positive integer greater than or equal to 50. A uniform random number generator is used to generate N different initial value matrices for charging and discharging power allocation. The rows of each initial value matrix for charging and discharging power allocation correspond to the number of buildings, and the columns correspond to the number of scheduling time windows. N sets of initial value matrices for charging and discharging power allocation are input into the dynamic game collaborative optimization model, and multiple rounds of iterative solution are executed in parallel. The number of iterations and the final Nash equilibrium solution after convergence are recorded for each set of initial value matrices. The Nash equilibrium solutions corresponding to the initial value matrix whose iteration convergence number exceeds the preset maximum iteration number are removed, and the remaining M Nash equilibrium solutions are retained, where M is less than or equal to N.

10. A method for collaborative optimization of energy storage resources for building complexes in a park, as described in claim 2, is characterized in that, The probabilistic prediction model is a deep learning model constructed based on a hybrid of long short-term memory neural networks and attention mechanisms.

Images