Energy storage system regulation and control method and device based on fusion terminal and terminal equipment

Abstract

The invention provides an energy storage system regulation and control method and device based on a fusion terminal and terminal equipment, and relates to the technical field of power distribution Internet of Things, and the method comprises the steps: obtaining an environment parameter and a historical load of a target station area in a specified time period through a target fusion terminal, and taking the environment parameter and the historical load as input, obtaining a predicted load of the target transformer area through a pre-trained load prediction model; obtaining time-of-use electricity price data, generating a day-ahead scheduling plan of the target transformer area based on the predicted load and the time-of-use electricity price data, and sending the day-ahead scheduling plan to the target fusion terminal; wherein the load prediction model is obtained by training a preset deep learning model based on an improved parameter optimization algorithm according to different historical loads of different transformer areas under different environmental parameters; the improved parameter optimization algorithm is used for updating model parameters of the deep learning model by calculating a first-order moment estimated value and a second-order moment estimated value of an iterative gradient. The method can effectively improve the prediction precision of the model on the transformer area load.

Description

Technical Field

[0001] This application relates to the field of power distribution Internet of Things (IoT) technology, specifically to a method for regulating an energy storage system based on a converged terminal, a device for regulating an energy storage system based on a converged terminal, a machine-readable storage medium, and a terminal device. Background Technology

[0002] With the transformation of the global energy structure, the penetration rate of renewable energy sources such as photovoltaics and wind power continues to rise. However, the output of these energy sources is affected by natural conditions, exhibiting significant intermittency, volatility, and randomness. For example, a surge in daytime photovoltaic power generation may lead to local grid overcapacity, while power shortages may occur at night or during windless periods. This uncertainty puts enormous pressure on the real-time balance of the power grid, power quality, and safe and stable operation. Power systems often struggle to absorb or store these sudden energy surges in a timely manner, easily leading to system overload or low-load operation. Battery Energy Storage Systems (BESS) are considered a core technological means to address these challenges. BESS can achieve energy transfer over time (i.e., "peak shaving and valley filling"), charging during periods of low grid load and releasing during peak loads, thereby smoothing the load curve and achieving load balancing. In addition, it can provide ancillary services such as rapid frequency response and voltage support, making it a key asset for improving power quality and grid resilience. In order to efficiently manage and dispatch the increasing number of distributed energy storage units, the architecture of new power systems is also constantly evolving. At present, the "cloud-edge-terminal" collaborative architecture, as an advanced cyber-physical fusion paradigm, is being widely used in the integrated regulation and control of "source, grid, load and storage".

[0003] Traditional energy storage system control often employs a centralized architecture, where all terminal data is aggregated to a cloud-based central control center for unified processing and decision-making. This model suffers from drawbacks such as high latency and low real-time performance, enormous communication and computational burdens, and reliability and privacy risks. Furthermore, in the absence of effective coordination mechanisms, many distributed energy storage systems often operate independently, resulting in relatively simplistic and lagging control strategies that lack sufficient intelligence and foresight, leading to a "whole is less than the sum of its parts" dilemma. This results in low resource utilization, and due to differences in manufacturing processes, operating temperatures, and charge / discharge histories, the state of health (SOH) of individual cells or battery modules gradually diverges. Under traditional balancing strategies based on state of charge (SOC), modules with lower SOH will reach their charge / discharge cutoff conditions prematurely, limiting the overall usable capacity and cycle life of the battery pack and exacerbating the "weakest link" effect. Therefore, a novel energy storage system control method is urgently needed to address the challenges of coordinated control, improve the overall operating efficiency of the energy storage system, extend battery life, achieve a better balance between lifespan and efficiency, enhance the grid's capacity to accommodate renewable energy, improve power supply reliability, and enhance the user's energy experience. Summary of the Invention

[0004] The purpose of this application is to provide a method for regulating an energy storage system based on a fusion terminal, a device for regulating an energy storage system based on a fusion terminal, a machine-readable storage medium, and a terminal device to solve the above-mentioned problems.

[0005] To achieve the above objectives, the first aspect of this application provides a method for regulating an energy storage system based on a converged terminal, applied to a cloud platform, comprising: The target fusion terminal obtains the environmental parameters and historical load of the target area within a specified time period. Using the environmental parameters and historical load as input, the predicted load of the target area is obtained through a pre-trained load prediction model. Acquire time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and the time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal; The load prediction model is obtained by training a preset deep learning model based on different historical loads of different transformer areas under different environmental parameters and an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model by calculating the first moment estimate and the second moment estimate of the iterative gradient during the training process of the deep learning model.

[0006] Optionally, the environmental parameters include temperature, humidity, air pressure, and wind speed; the preset deep learning model includes multiple stacked feature extraction network layers and a backpropagation network layer; training the preset deep learning model based on an improved parameter optimization algorithm includes: A training sample set is constructed using the temperature, humidity, air pressure, wind speed of different transformer areas at different historical periods, as well as the historical load of each transformer area at each corresponding historical period. The training sample set is then divided into multiple training sample subsets. The network parameters of the feature extraction network layer are initialized, and the preset deep learning model is trained sequentially with each subset of training samples as input. The network parameters of the feature extraction network layer and the backpropagation network layer are updated through the improved parameter optimization algorithm until the convergence condition is met. The mapping relationship between the input and output of the preset deep learning model is determined, and the load prediction model is obtained.

[0007] Optionally, the feature extraction network layer includes at least one visible layer and one hidden layer; the network parameters of the feature extraction network layer are updated using the improved parameter optimization algorithm, including: Construct a momentum vector and determine the initial value of the momentum vector. The momentum vector includes the weight coefficients between the visible and hidden layers of the corresponding feature extraction network layer, the activation state of each node in the visible layer, and the activation state of each node in the hidden layer. An energy function representing the working state of the visible layer and the hidden layer is constructed based on the momentum vector; Based on the energy function, a joint probability distribution function is established to indicate that each node in the visible layer and the hidden layer is simultaneously in a specific activation state. Based on the joint probability distribution function, a conditional probability function is established to characterize the activation state of each node in the visible layer and the hidden layer. With the goal of maximizing the joint probability, an update model for the momentum vector is constructed based on the first-order moment estimate and the second-order moment estimate of the iterative gradient. The momentum vector is updated until the convergence condition is met, and the current momentum vector is used as the network parameter of the preset deep learning model.

[0008] Optionally, the momentum vector update model includes:

[0009]

[0010]

[0011]

[0012]

[0013]

[0014] in, The gradient vector, Let θ be the objective function, and θ represent the momentum vector. The attenuation constant is For learning efficiency, s is the number of iterations, and i represents the i-th feature extraction network layer. This represents the gradient operator at the s-th iteration of the i-th feature extraction network layer. and Let these represent the first and second moment estimates of the gradient of the i-th feature extraction network layer at the s-th iteration, respectively. This represents the gradient operator for the i-th feature extraction network layer during the (s-1)-th iteration. This represents the momentum vector of the i-th feature extraction network layer during the s-th iteration. Let represent the momentum vector of the i-th feature extraction network layer at the (s+1)-th iteration. and Let represent the first-order moment estimate correction value and the second-order moment estimate correction value of the gradient at the s-th iteration of the i-th feature extraction network layer, respectively. , and To correct the parameters.

[0015] Optionally, the network parameters of the backpropagation network layer include the weight coefficients between the input neurons and output neurons of the backpropagation network layer; updating the network parameters of the backpropagation network layer using the improved parameter optimization algorithm includes: A loss function based on regularization terms is constructed based on the predicted and input values ​​of the preset deep learning model, and a penalty function is constructed based on the weight coefficients between the input and output neurons of the backpropagation network layer. The target loss function is constructed based on the regularization-based loss function and the penalty function. The weight coefficients between the input and output neurons of the backpropagation network layer are updated based on the target loss function until the convergence condition is met. The current weight coefficients between the input and output neurons of the backpropagation network layer are then used as the network parameters of the backpropagation network layer.

[0016] Optionally, the target loss function includes:

[0017]

[0018] in, Let be the target loss function. and These are the actual value and the predicted value, respectively. Let m be the input feature vector of the x-th sample, and m be the number of samples. Let be the penalty function. The proportionality coefficient of the penalty function, For the loss function based on the regularization term, These are the weight coefficients between the input and output neurons of the backpropagation network layer.

[0019] A second aspect of this application provides a method for regulating an energy storage system based on a converged terminal, applied to a converged terminal, comprising: Send the environmental parameters and historical load of the target area within the specified time period to the cloud platform; The system receives the day-ahead scheduling plan for the target distribution area generated by the cloud platform based on the environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted load for the target distribution area during different day-ahead scheduling periods and the charging / discharging status of the energy storage units during those periods. Using environmental parameters within the specified time period as input, the predicted power output of the target power distribution area at different times is obtained through the power output prediction model. The real-time power output of the target power distribution area is obtained. Based on the error between the real-time power output of the target power distribution area and the predicted power output at the corresponding time, the energy storage unit of the target power distribution area is controlled to charge and discharge, so that the real-time power output of the target power distribution area matches the corresponding predicted power output. The power output prediction model is obtained by training a preset deep learning model on a cloud platform based on historical environmental parameters and historical power output of different power distribution areas.

[0020] Optionally, controlling the charging and discharging of the energy storage units in the target power distribution area based on the error between the real-time output power and the predicted output power at the corresponding time includes: If the real-time output power of the target power station is lower than the predicted output power of the target power station at the current time, the output power difference between the real-time output power and the predicted output power is determined. The energy storage unit is controlled to discharge based on the output power difference until the real-time output power of the target power station matches the predicted output power at the current sampling time, at which point the energy storage unit stops discharging.

[0021] Optionally, the energy storage unit includes at least one battery module, and controlling the energy storage unit in the target area to charge and discharge includes: Obtain the SOH value, which characterizes the health status of each battery module, and the cumulative number of cycles for each battery module; Sort the battery modules by SOH value from high to low, or sort the battery modules by cumulative cycle count from low to high; To determine the target discharge power, during the discharge process of the energy storage unit, target battery modules are determined sequentially from high to low according to the sorting order of each battery module until the total discharge power of each target battery module is not less than the target discharge power.

[0022] Optionally, controlling the energy storage unit of the target substation area to charge and discharge further includes: Energy storage units with SOH values ​​below a preset threshold are identified as target energy storage units. During the charging process of the energy storage unit, the charging current of the target energy storage unit is controlled to be no greater than a set charging current, which is less than the rated charging current of the energy storage unit.

[0023] A third aspect of this application provides a control device for an energy storage system based on a converged terminal, which applies the above-described control method for an energy storage system based on a converged terminal. The device includes: The load forecasting module is configured to obtain the environmental parameters and historical load of the target transformer area within a specified time period through the target fusion terminal, and obtain the predicted load of the target transformer area by using the environmental parameters and the historical load as input and a pre-trained load forecasting model. The plan generation module is configured to acquire time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and the time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal; wherein, the load prediction model is obtained by training a preset deep learning model based on different historical loads of different distribution areas under different environmental parameters, using an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model during the training process by calculating the first moment estimate and second moment estimate of the iterative gradient.

[0024] In a fourth aspect, this application provides a control device for an energy storage system based on a converged terminal, which applies the above-described control method for an energy storage system based on a converged terminal. The device includes: The data transmission module is configured to send the environmental parameters and historical load of the target area within a specified time period to the cloud platform; The data receiving module is configured to receive the day-ahead scheduling plan of the target distribution area generated by the cloud platform based on the environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted load of the target distribution area in different day-ahead scheduling periods and the charging and discharging status of the energy storage unit in different day-ahead scheduling periods. The control module is configured to take environmental parameters within the specified time period as input, obtain the predicted power output of the target power distribution area at different times through a power output prediction model, acquire the real-time power output of the target power distribution area, and control the charging and discharging of the energy storage unit of the target power distribution area based on the error between the real-time power output of the target power distribution area and the predicted power output at the corresponding time, so that the real-time power output of the target power distribution area matches the corresponding predicted power output; wherein, the power output prediction model is obtained by the cloud platform after training a preset deep learning model based on the historical environmental parameters and historical power output of different power distribution areas.

[0025] In a fifth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform either the above-described method for regulating an energy storage system based on a converged terminal applied to a cloud platform, or the above-described method for regulating an energy storage system based on a converged terminal applied to a converged terminal.

[0026] In a sixth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for regulating an energy storage system based on a converged terminal applied to a cloud platform, or implements the steps of the above-described method for regulating an energy storage system based on a converged terminal applied to a converged terminal.

[0027] This application utilizes an improved parameter optimization algorithm. By calculating the first and second moment estimates of the iterative gradient during the iteration process of a deep learning model, the descent gradient can be estimated. This effectively accelerates the convergence process, improves model training efficiency and effectiveness, and enhances the model's accuracy in predicting transformer load. Consequently, it enables the formulation of smoother, more orderly, and planned charging and discharging strategies based on load prediction results.

[0028] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A flowchart illustrating a preferred embodiment of this application for a method of regulating an energy storage system based on a fusion terminal; Figure 2 This is a schematic diagram of a cloud-edge-device architecture provided in a preferred embodiment of this application; Figure 3 This is a schematic diagram of the DBN model structure provided in a preferred embodiment of this application; Figure 4 This is a schematic diagram of the load prediction model training process provided in a preferred embodiment of this application; Figure 5 A schematic diagram of another energy storage system control method based on a fusion terminal provided in a preferred embodiment of this application; Figure 6 This is a schematic diagram of battery module equalization scheduling provided in a preferred embodiment of this application; Figure 7 A schematic block diagram of an energy storage system control device based on a fusion terminal, provided as a preferred embodiment of this application; Figure 8 A schematic block diagram of another energy storage system control device based on a fusion terminal, provided as a preferred embodiment of this application; Figure 9 A schematic diagram of a terminal device provided for a preferred embodiment of this application.

[0030] Description of Reference Numerals 10 - Terminal device, 100 - Processor, 101 - Memory, 102 - Computer program. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0032] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0033] Existing technologies suffer from several drawbacks, including uneven battery aging leading to significant differences in cycle life between battery packs, severe battery capacity loss resulting in a continuous decline in usable capacity, the inability to promptly remove aged batteries from the charge / discharge sequence affecting operational safety, a lack of battery pack load balancing control causing capacity loss, and the inability to predict whether charging power will meet future power supply duration or load demands. Furthermore, traditional methods often lack a unified coordination mechanism and information sharing platform, and some existing balancing methods may have slow response times, making it difficult to adapt to rapidly changing grid conditions and load demands. Moreover, their control strategies are often simplistic, primarily passive, lacking the ability to predict future conditions and proactively adjust, thus requiring improvements in balancing accuracy and efficiency.

[0034] To solve the above problems, such as Figure 1 As shown, the first aspect of this application provides a method for regulating an energy storage system based on a converged terminal, applied to a cloud platform, comprising: S100. Obtain the environmental parameters and historical load of the target transformer area within a specified time period through the target fusion terminal. Using the environmental parameters and historical load as input, obtain the predicted load of the target transformer area through a pre-trained load prediction model. S200: Obtain time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal; wherein, the load prediction model is obtained by training a preset deep learning model based on different historical loads of different distribution areas under different environmental parameters, using an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model by calculating the first moment estimate and second moment estimate of the iterative gradient during the training process of the deep learning model.

[0035] Thus, by using an improved parameter optimization algorithm, the descent gradient is estimated by calculating the first and second moment estimates of the iterative gradient during the iteration process of the deep learning model. This effectively accelerates the convergence process, improves model training efficiency and effectiveness, and enhances the model's prediction accuracy of the load in the transformer area. Consequently, a smoother, more orderly, and planned charging and discharging strategy can be formulated based on the load prediction results.

[0036] like Figure 2As shown, the energy storage control system of this application is built on a cloud-edge-device architecture. The cloud platform, serving as the system's decision-making center and data brain, is deployed on a remote server. Its core functions include: long-term forecasting and planning, for example, by running heavyweight deep learning models such as DBN (Deep Belief Network) models to accurately predict the regional load and renewable energy output (photovoltaic and wind power) of a specified area, such as each transformer substation for the next 24 hours or even longer; global economic dispatching, for example, by formulating a globally optimal economic dispatching plan (day-ahead plan) on a daily basis based on forecast results, real-time electricity prices, grid demand, and battery aging cost models; and big data storage and analysis, aggregating and permanently storing key historical data uploaded by each edge node for model iteration training, system health status assessment, and long-term operation strategy optimization.

[0037] A converged terminal is a high-performance edge computing gateway that can be deployed in distribution substations, energy storage containers, or user-side substations. Its main functions are: communicating with the cloud platform via 4G / 5G wireless networks or fiber optics to receive day-ahead scheduling plans from the cloud, and selectively and periodically uploading local key aggregated data (such as total load of the distribution area and average state of charge (SOH) of battery clusters) to avoid raw data overload. The converged terminal can perform high-frequency (e.g., 15-minute or shorter cycles) rolling corrections and closed-loop control locally, directly collecting real-time data such as local load and energy storage status, comparing it with day-ahead plans, and quickly responding to unplanned fluctuations. It achieves high-speed and reliable data exchange with energy storage BMS, inverters, and sensors via industrial buses such as CAN bus and RS485, and communicates with the cloud using lightweight IoT protocols such as MQTT, enabling seamless integration of heterogeneous networks and protocols.

[0038] The energy storage unit consists of multiple modular battery modules (packs), each with an independent built-in battery management system (BMS) responsible for monitoring the voltage and temperature of individual cells and providing basic protection. The battery pack has communication capabilities; the modular energy storage batteries are inserted into the unit's expansion slots, supporting capacity expansion and replacement. The PCS has the ability to communicate with the modular batteries (battery packs) and can connect to the smart terminal in the distribution area for remote control and the distribution of control strategies.

[0039] The data acquisition unit includes smart meters, voltage / current transformers, temperature sensors, etc., and is used to provide accurate raw data to the edge layer and the cloud.

[0040] In step S100, the environmental parameters of this application include, but are not limited to, temperature, humidity, air pressure, and wind speed. In this application, the preset deep learning model includes multiple stacked feature extraction network layers and a backpropagation network layer. For example, the preset deep learning model can be a DBN model, such as... Figure 3As shown, due to the nonlinearity, randomness, and chaos of power load data, the DBN model constructed in this application includes multiple stacked Restricted Boltzmann Machines (RBMs) and a Backpropagation Neural Network (BPNN) with one hidden layer. Each RBM is a feature extraction network layer, and the BPNN is the backpropagation network layer. The DBN model is used to input data (such as temperature, humidity, air pressure, wind speed, weekdays, etc.) in an unsupervised manner, extract the hidden features of the input data, and use the BPNN with one hidden layer to perform nonlinear fitting of the feature quantities in a supervised manner. It is understandable that the DBN model of the fusion terminal uses the same type of input variables and network structure; the difference lies in the final weight coefficients obtained during training.

[0041] In this application, a pre-defined deep learning model is trained based on an improved parameter optimization algorithm. This includes: constructing a training sample set using temperature, humidity, air pressure, wind speed, and historical load data for different power distribution areas at different historical time periods, and dividing the training sample set into multiple training sample subsets; initializing the network parameters of the feature extraction network layer; sequentially training the pre-defined deep learning model using each training sample subset as input; updating the network parameters of the feature extraction network layer and the backpropagation network layer using the improved parameter optimization algorithm until convergence is met; determining the mapping relationship between the input and output of the pre-defined deep learning model; and obtaining the load prediction model. For example, temperature, humidity, air pressure, wind speed, and historical load data for different power distribution areas at different historical time periods are obtained, and historical data from each historical time period are used as training samples to construct a training sample set. Multiple training sample data in the training sample set are then divided into G groups of mini-batch sample data. When training the DBN model, in each iteration, a set of training sample data is obtained from the training sample set in sorted order to train the DBN model.

[0042] In this application, the restricted Boltzmann machine comprises two layers, the first of which is the visible layer. The first layer is used to receive external input variables, such as environmental and time-series variables, or the output from the second-level restricted Boltzmann machine. The second layer is a hidden layer. This is used to extract feature information contained in the input layer data. A fully connected structure with weighted coefficients is used between the visible and hidden layers, while units within the same layer are not connected to each other. For example... Figure 4 As shown, in this application, the network parameters of the feature extraction network layer are updated using an improved parameter optimization algorithm, including: S10. Construct the momentum vector and determine its initial value. The momentum vector includes the weight coefficients between the visible and hidden layers of the corresponding feature extraction network layer, the activation states of each node in the visible layer, and the activation states of each node in the hidden layer. Let the first... The input and output components of each RBM are respectively and The set of weight coefficients between the visible layer and the hidden layer is ,in, The activation states of each node (visible element) in the visible layer and each node (hidden element) in the hidden layer satisfy the following conditions: Therefore, for an RBM, one can use... This is used to represent its network parameters.

[0043] S20. Construct an energy function based on momentum vector to represent the working state of the visible and hidden layers.

[0044] Specifically, this application uses the following energy function to measure the operating state of the RBM:

[0045] in, Let represent the energy of the i-th RBM with network parameters θ. Each RBM has n visible layer nodes and m hidden layer nodes. This represents the connection weight between the visible layer node p and the hidden layer node q of the i-th RBM. This indicates the activation state of the p-th visible layer node of the i-th RBM. This represents the activation state of the hidden layer node of the i-th RBM. This represents the state value of the p-th visible layer node of the i-th RBM. This represents the state value of the q-th hidden layer node of the i-th RBM.

[0046] S30. Based on the energy function, establish a joint probability distribution function for each node in the visible and hidden layers simultaneously being in a specific activation state, and based on the joint probability distribution function, establish a conditional probability function characterizing the activation state of each node in the visible and hidden layers. Specifically, when the network parameters are determined, the joint probability distribution of (v, h) can be obtained. The joint probability distribution function of this application is as follows:

[0047] in, The normalization factor, i.e., the first The energy function of all possible states of an RBM, since the nodes in the same layer are not connected to each other, their activation states are conditionally independent. Therefore, the energy function of the RBM is... The visible unit and the first One hidden unit is activated, that is , The probability of equaling 1 can be expressed by the conditional probability function as:

[0048]

[0049] in, It is a sigmoid activation function, and .

[0050] S40. With the goal of maximizing the joint probability, an update model for the momentum vector is constructed based on the estimated first and second moments of the iterative gradient. This model updates the momentum vector until the convergence condition is met, and the current momentum vector is used as the network parameter of the preset deep learning model. It is understandable that the value of the momentum vector directly determines the magnitude of the energy function, thus affecting the calculation results of the joint probability distribution and conditional probability (activation state probability). During the momentum vector update process, by adjusting parameters such as weights and biases, the value of the energy function is changed, leading to a change in the joint probability distribution of the visible and hidden layers. Ultimately, this adjusts the activation state probability of the units (such as the probability of a hidden layer node being activated). The learning objective of RBM is to maximize the joint probability of the input data by optimizing the network parameters, i.e., the probability of the joint distribution of the visible and hidden layers given the input. Conditional probability can be used to approximate the gradient update by providing a computationally comprehensible iterative process for optimizing the joint probability. For example, the gradient operator can be estimated using conditional probability. The process of estimating the gradient operator using conditional probability is existing technology and is not limited here.

[0051] In traditional DBN, the gradient update rule is as follows:

[0052]

[0053]

[0054] Therefore, the momentum update rule can be expressed as:

[0055] in, The gradient vector, To obtain the objective function value, in traditional DBN, the training objective is to maximize the likelihood function of the observed data (visible layer states), that is, to maximize the joint probability of the input data by optimizing the network parameters. Let be the momentum vector with an initial value of 0. The attenuation constant is For learning efficiency, s is the number of iterations.

[0056] Because the decay constant and learning rate are fixed constants in traditional gradient update rules, and momentum updates are only related to the momentum and gradient direction of the previous iteration, the DBN training method with fixed parameters may not achieve good prediction accuracy within a limited training time.

[0057] Understandably, in this application, there is a strong correlation between load forecasting and battery module aging. The accuracy of load forecasting directly determines the charging and discharging strategy of the energy storage system, which is the most critical external factor affecting battery aging. High-precision load forecasting enables the formulation of smooth, orderly, and planned charging and discharging strategies. This allows the energy storage system to avoid frequent deep charging and discharging and high-rate charging and discharging, which are the main causes of accelerated battery capacity decay and increased internal resistance. Inaccurate load forecasting forces the energy storage system to frequently and drastically respond to forecast errors, performing emergency and disordered charging and discharging to maintain grid balance. This high-intensity, random operating mode significantly increases the number of battery cycles, deepens the average depth of discharge, and may cause local overheating due to rapid charging and discharging, thereby drastically accelerating the decay of the battery's state of harmonic equilibrium (SOH) and shortening its economic life.

[0058] To address the aforementioned issues, this application employs first-order moment estimation and second-order moment estimation in gradient descent to improve the momentum vector update process. The improved momentum vector update model can be expressed as:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Based on the above, the updated momentum model can be transformed into:

[0065] in, The gradient vector, Let θ be the objective function, and θ represent the momentum vector. The attenuation constant is For learning efficiency, s is the number of iterations, and i represents the i-th feature extraction network layer. This represents the gradient operator at the s-th iteration of the i-th feature extraction network layer. and Let these represent the first and second moment estimates of the gradient of the i-th feature extraction network layer at the s-th iteration, respectively. This represents the gradient operator for the i-th feature extraction network layer during the (s-1)-th iteration. This represents the momentum vector of the i-th feature extraction network layer during the s-th iteration. Let represent the momentum vector of the i-th feature extraction network layer at the (s+1)-th iteration. and Let represent the first-order moment estimate correction value and the second-order moment estimate correction value of the gradient at the s-th iteration of the i-th feature extraction network layer, respectively. , and To correct the parameters. Understandably, the first-order moment estimate can be the mean of the gradient, and the second-order moment estimate can be the mean of the squared gradients. This application uses a BPNN with one hidden layer, and based on the data features extracted by the DBN, utilizes the nonlinear fitting capability of the BPNN to predict the load. Assume there are N RBMs in the DBN, and the final data features extracted by the DBN are... Let the BPNN output be the load forecast value, denoted as y, then In this application, the network parameters of the backpropagation network layer include the weight coefficients between the input neurons and the output neurons of the backpropagation network layer. Specifically, Here, K represents the weight coefficients between the input and output neurons in a BPNN, and K is the number of neurons in the hidden layer. This represents the q-th feature extracted by DBN. To reduce the weight changes between input and output neurons during each training iteration and make the weight changes smoother, this application adds a regularization term to the loss function. When training on samples with a large number of noisy features, the model will not overfit due to these samples, thus better balancing the relationship between training error and model complexity. In this application, the network parameters of the backpropagation network layer are updated through an improved parameter optimization algorithm. This includes: constructing a regularization-based loss function based on the predicted and input values ​​of a pre-defined deep learning model, and constructing a penalty function based on the weight coefficients between the input and output neurons of the backpropagation network layer; constructing a target loss function based on the regularization-based loss function and the penalty function; updating the weight coefficients between the input and output neurons of the backpropagation network layer based on the target loss function until the convergence condition is met, using the current weight coefficients between the input and output neurons of the backpropagation network layer as the network parameters of the backpropagation network layer. The convergence condition can be that the error between the predicted and actual values ​​is less than a threshold, or that the maximum number of iterations is reached. Specifically, the target loss function of this application includes:

[0066]

[0067] in, Let be the target loss function. and These are the actual value and the predicted value, respectively. Let m be the input feature vector of the x-th sample, and m be the number of samples. Let be the penalty function. The proportionality coefficient of the penalty function, For the loss function based on the regularization term, These are the weight coefficients between the input and output neurons of the backpropagation network layer.

[0068] In this application, the cloud platform also trains the DBN model based on historical environmental parameters and historical power output of different distribution areas sent by the fusion terminal to obtain a power output prediction model for predicting the power output of new energy in the distribution area. The training method for the power output prediction model is the same as described above. After training, the model is sent to the fusion terminal. In step S200, the cloud platform generates a preliminary hourly charging and discharging plan based on the predicted load generated by the load prediction model, combined with information such as time-of-use pricing and grid peak-shaving demand, with a 24-hour cycle and aiming at minimizing the total system operating cost. This plan is then sent to the fusion terminal. The plan defines the charging, discharging, or standby state that the energy storage system should be in each hour, as well as the approximate power level. The fusion terminal can use the power output prediction model to make real-time corrections and short-term predictions of the charging and discharging plan.

[0069] like Figure 5 As shown, in a second aspect, this application provides a method for regulating an energy storage system based on a converged terminal, applied to a converged terminal, comprising: S300 sends the environmental parameters and historical load of the target area within a specified time period to the cloud platform; S400 receives the day-ahead scheduling plan for the target distribution area generated by the cloud platform based on environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted load of the target distribution area during different day-ahead scheduling periods and the charging and discharging status of the energy storage unit during different day-ahead scheduling periods. S500: Using environmental parameters within a specified time period as input, the predicted power output of the target transformer area at different times is obtained through the power output prediction model. The real-time power output of the target transformer area is obtained. Based on the error between the real-time power output of the target transformer area and the predicted power output of the corresponding day-ahead scheduling period, the energy storage unit of the target transformer area is controlled to charge and discharge, so that the real-time power output of the target transformer area matches the corresponding predicted power output. The power output prediction model is obtained by the cloud platform after training a preset deep learning model based on the historical environmental parameters and historical power output of different transformer areas.

[0070] In step S500, the energy storage unit of the target distribution area is charged and discharged based on the error between the real-time output power of the target distribution area and the predicted load of the corresponding day-ahead scheduling period. This includes: if the real-time output power of the target distribution area at the current sampling time is lower than the predicted output power of the target distribution area at the current sampling time, the output power difference between the real-time output power and the predicted output power is determined; the energy storage unit is discharged based on the output power difference until the real-time output power of the target distribution area matches the predicted output power at the current sampling time, and the energy storage unit is controlled to stop discharging.

[0071] Since the day-ahead plan is a macro-level guide based on forecasts, and actual situations often deviate from the forecast, this application uses a fusion terminal to make rolling corrections to the day-ahead plan at a shorter cycle (e.g., 15 minutes), forming a fast-response closed-loop control.

[0072] For example, suppose an industrial and commercial park is equipped with an energy storage system, and its time-of-use electricity pricing is as follows: off-peak hours (00:00-06:00) 0.3 yuan / kWh, normal hours (06:00-18:00) 0.7 yuan / kWh, and peak hours (18:00-22:00) 1.2 yuan / kWh. The cloud platform predicts a peak electricity demand of 500kW lasting for 2 hours from 19:00 to 21:00 the following day. Therefore, the day-ahead plan formulated by the cloud platform is as follows: Charging plan: From 02:00 to 04:00 the next day (off-peak electricity price period), the energy storage system will be instructed to charge from the grid at a power of 500kW, for a total of 1000kWh. Discharging plan: From 19:00 to 21:00 the next day (peak electricity price period), the energy storage system will be instructed to discharge to the park load at a power of 500kW to meet peak demand and avoid purchasing high-priced electricity from the grid. During other times, the energy storage system will be in standby mode.

[0073] Suppose the energy storage system is scheduled to be on standby from 2:00 PM to 4:00 PM, but at 2:30 PM, the weather suddenly changes, with large dark clouds causing the park's photovoltaic output to plummet from the planned 600kW to 100kW, a reduction of 500kW. Without energy storage intervention, the park would instantly draw an extra 500kW of power from the grid, potentially leading to transformer overload or high demand charges.

[0074] The integrated terminal in this application can detect this deviation within seconds by monitoring the data from the photovoltaic inverter and the power grid interface meter in real time. Upon detecting a decrease in photovoltaic output in the park, the integrated terminal immediately makes a decision, deviating from the day-ahead plan, and instructs the energy storage system to discharge at a power of 500kW to precisely compensate for the photovoltaic output shortfall and maintain the stability of the park's total load. For example, when the weather improves at 15:00 and photovoltaic output recovers, the integrated terminal will instruct the energy storage system to stop discharging and return to standby mode. The entire process is completed autonomously by the integrated terminal locally, without cloud intervention, ensuring timely response.

[0075] In this application, the energy storage unit includes at least one battery module. The fusion terminal collects and stores key state data from the BMS of each battery module (Pack) at a high frequency (e.g., at the second level). This data includes not only real-time SOC and temperature, but also indicators reflecting long-term aging trends, such as cumulative charge-discharge cycle count and cumulative charge-discharge ampere-hours. Based on parameters such as internal resistance and capacity, the state of health (SOH) of each module is estimated. When performing charge-discharge tasks, the fusion terminal no longer treats all battery modules as homogeneous resources, but implements differentiated dynamic scheduling strategies based on their key state data. Specifically, controlling the energy storage units in the target area to charge and discharge includes: acquiring the SOH value and cumulative cycle count of each battery module; sorting the battery modules from high to low SOH value, or from low to high cumulative cycle count; determining the target discharge power. During the discharge process of the energy storage unit, the target battery modules are determined sequentially from high to low according to the sorting order of the battery modules until the total discharge power of the target battery modules is not less than the target discharge power.

[0076] For example, such as Figure 6 As shown, an energy storage container contains three parallel battery clusters, each consisting of 10 battery modules (packs) connected in series. The entire container needs to be discharged at a power of 300kW, and the fusion terminal monitors the health status of the 10 modules within one of the battery clusters as follows: Module AG (7 units): SOH ≈ 95%, cumulative number of cycles ≈ 800.

[0077] Modules H and I (2 units): SOH ≈ 94%, cumulative number of cycles ≈ 850.

[0078] Module J (1 unit): SOH ≈ 88%, cumulative cycle count ≈ 1200 times. This is determined to be the "weakest link" module.

[0079] The traditional control strategy is to distribute power evenly across all modules, with each module discharging 30kW. This will cause module J to age much faster than other modules, and its capacity will soon severely degrade, leading to a decline in the performance of the entire cluster.

[0080] The SOH balancing strategy of this application is as follows: Discharge task allocation: The fusion terminal identifies module J as having the worst health condition. When allocating a 300kW discharge task, it prioritizes healthy modules. For example, each of the nine instruction modules AI discharges approximately 33.3kW, while instruction module J is in standby or low-power discharge mode. In this way, while completing system tasks, module J gets a "rest," slowing down its aging process.

[0081] In this application, controlling the energy storage units in the target area to charge and discharge also includes: determining the energy storage units with SOH values ​​lower than a preset threshold as target energy storage units, and controlling the charging current of the target energy storage unit to be no greater than a set charging current during the charging process of the energy storage unit, wherein the set charging current is less than the rated charging current of the energy storage unit.

[0082] Specifically, the charging task allocation is as follows: During charging, if there is sufficient time, priority can be given to charging module J with a lower SOH, or all modules can be charged simultaneously, but a smaller charging current (e.g., 0.25C) can be set for module J, while other modules use normal current (e.g., 0.5C). By reducing the charging current of low-health energy storage modules, further damage to aging batteries can be effectively reduced.

[0083] In this application, the integrated terminal also performs fault diagnosis and aging warning for the energy storage system. The integrated terminal continuously monitors the operating data of all battery modules and establishes a normal behavior baseline for each module based on this data. When the operating parameters of a module (such as SOH, temperature rise, voltage curve, etc.) deviate from its normal baseline or show significant abnormalities compared with other modules, the system determines whether there is an aging battery module and triggers an early warning. For example, if the integrated terminal finds through long-term monitoring and calculation that the estimated SOH value of a battery module is lower than the preset maintenance threshold, such as the industry-standard 80%, then it determines that the module is aging; alternatively, it can also determine the aging of battery modules through abnormal temperature rises or abnormal voltage curves, such as when the voltage platform of the module drops too quickly during charging and discharging, or when the voltage difference (the difference between the highest and lowest single-cell voltages within the same module) continuously exceeds the standard.

[0084] If the battery module is determined to be aging, the converged terminal will mark it as "needing maintenance" or "faulty." In subsequent scheduling, its usage priority will be automatically reduced or it will be completely isolated. The converged terminal will then control the entire machine to remove the aging battery module from the charging / discharging sequence, add redundant battery modules to the charging / discharging sequence, and send detailed warning information to the cloud-based operations and maintenance platform to guide maintenance personnel in accurate inspection or replacement. Otherwise, the above-mentioned balancing strategy will be returned.

[0085] like Figure 7 As shown, in a third aspect of this application, a control device for an energy storage system based on a converged terminal is provided. Applying the aforementioned control method for an energy storage system based on a converged terminal, the device includes: The load forecasting module is configured to obtain the environmental parameters and historical load of the target transformer area within a specified time period through the target fusion terminal, and obtain the predicted load of the target transformer area by taking the environmental parameters and historical load as input and through a pre-trained load forecasting model. The plan generation module is configured to acquire time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal. The load prediction model is obtained by training a preset deep learning model based on different historical loads of the target distribution area under different environmental parameters and an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model by calculating the first moment estimate and second moment estimate of the iterative gradient during the training process of the deep learning model.

[0086] like Figure 8 As shown, in a fourth aspect of this application, a control device for an energy storage system based on a converged terminal is provided. Applying the aforementioned control method for an energy storage system based on a converged terminal, the device includes: The data transmission module is configured to send the environmental parameters and historical load of the target area within a specified time period to the cloud platform; The data receiving module is configured to receive the day-ahead scheduling plan of the target transformer area generated by the cloud platform based on environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted output power of the target transformer area during different day-ahead scheduling periods and the charging and discharging status of the energy storage unit during different day-ahead scheduling periods. The control module is configured to acquire the real-time output power of the target power distribution area, and control the energy storage units of the target power distribution area to charge and discharge based on the error between the real-time output power of the target power distribution area and the predicted output power of the corresponding day-ahead scheduling period, so as to match the real-time output power of the target power distribution area with the corresponding predicted output power.

[0087] In a fifth aspect, this application provides a machine-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform either the above-described method for regulating an energy storage system based on a converged terminal applied to a cloud platform, or the above-described method for regulating an energy storage system based on a converged terminal applied to a converged terminal.

[0088] In a sixth aspect, this application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for regulating an energy storage system based on a converged terminal applied to a cloud platform, or implements the steps of the above-described method for regulating an energy storage system based on a converged terminal applied to a converged terminal.

[0089] like Figure 9 The diagram shown is a schematic representation of a terminal device provided in an embodiment of this application. Figure 9 As shown, the terminal device 10 of this embodiment includes a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, it implements the steps in the above method embodiments. Alternatively, when the processor 100 executes the computer program 102, it implements the functions of each module / unit in the above device embodiments.

[0090] For example, computer program 102 may be divided into one or more modules / units, one or more of which are stored in memory 101 and executed by processor 100 to complete this application. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 102 in terminal device 10.

[0091] Terminal device 10 may be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. Terminal device 10 may include, but is not limited to, a processor 100 and a memory 101. Those skilled in the art will understand that... Figure 9 This is merely an example of terminal device 10 and does not constitute a limitation on terminal device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, terminal device may also include input / output devices, network access devices, buses, etc.

[0092] Processor 100 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0093] The memory 101 can be an internal storage unit of the terminal device 10, such as a hard disk or RAM of the terminal device 10. The memory 101 can also be an external storage device of the terminal device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the terminal device 10. Furthermore, the memory 101 can include both internal and external storage units of the terminal device 10. The memory 101 is used to store computer programs and other programs and data required by the terminal device 10. The memory 101 can also be used to temporarily store data that has been output or will be output.

[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0096] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for regulating an energy storage system based on a converged terminal, applied to a cloud platform, characterized in that, include: The target fusion terminal obtains the environmental parameters and historical load of the target area within a specified time period. Using the environmental parameters and historical load as input, the predicted load of the target area is obtained through a pre-trained load prediction model. Acquire time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and the time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal; The load prediction model is obtained by training a preset deep learning model based on different historical loads of different transformer areas under different environmental parameters and an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model by calculating the first moment estimate and the second moment estimate of the iterative gradient during the training process of the deep learning model.

2. The energy storage system control method based on a fusion terminal according to claim 1, characterized in that, The environmental parameters include temperature, humidity, air pressure, and wind speed; the preset deep learning model includes multiple stacked feature extraction network layers and a backpropagation network layer; training the preset deep learning model based on an improved parameter optimization algorithm includes: A training sample set is constructed using the temperature, humidity, air pressure, wind speed of different transformer areas at different historical periods, as well as the historical load of each transformer area at each corresponding historical period. The training sample set is then divided into multiple training sample subsets. The network parameters of the feature extraction network layer are initialized, and the preset deep learning model is trained sequentially with each subset of training samples as input. The network parameters of the feature extraction network layer and the backpropagation network layer are updated through the improved parameter optimization algorithm until the convergence condition is met. The mapping relationship between the input and output of the preset deep learning model is determined, and the load prediction model is obtained.

3. The energy storage system control method based on a fusion terminal according to claim 2, characterized in that, The feature extraction network layer includes at least one visible layer and one hidden layer; the network parameters of the feature extraction network layer are updated using the improved parameter optimization algorithm, including: Construct a momentum vector and determine the initial value of the momentum vector. The momentum vector includes the weight coefficients between the visible and hidden layers of the corresponding feature extraction network layer, the activation state of each node in the visible layer, and the activation state of each node in the hidden layer. An energy function representing the working state of the visible layer and the hidden layer is constructed based on the momentum vector; Based on the energy function, a joint probability distribution function is established to indicate that each node in the visible layer and the hidden layer is simultaneously in a specific activation state. Based on the joint probability distribution function, a conditional probability function is established to characterize the activation state of each node in the visible layer and the hidden layer. With the goal of maximizing the joint probability, an update model for the momentum vector is constructed based on the first-order moment estimate and the second-order moment estimate of the iterative gradient. The momentum vector is updated until the convergence condition is met, and the current momentum vector is used as the network parameter of the preset deep learning model.

4. The energy storage system control method based on a fusion terminal according to claim 3, characterized in that, The momentum vector update model includes: in, The gradient vector, Let θ be the objective function, and θ represent the momentum vector. The attenuation constant is For learning efficiency, s is the number of iterations, and i represents the i-th feature extraction network layer. This represents the gradient operator at the s-th iteration of the i-th feature extraction network layer. and Let these represent the first and second moment estimates of the gradient of the i-th feature extraction network layer at the s-th iteration, respectively. This represents the gradient operator for the i-th feature extraction network layer during the (s-1)-th iteration. This represents the momentum vector of the i-th feature extraction network layer during the s-th iteration. Let represent the momentum vector of the i-th feature extraction network layer at the (s+1)-th iteration. and Let represent the first-order moment estimate correction value and the second-order moment estimate correction value of the gradient at the s-th iteration of the i-th feature extraction network layer, respectively. , and To correct the parameters.

5. The energy storage system control method based on a fusion terminal according to claim 2, characterized in that, The network parameters of the backpropagation network layer include the weight coefficients between the input neurons and the output neurons of the backpropagation network layer. The network parameters of the backpropagation network layer are updated using the improved parameter optimization algorithm, including: A loss function based on regularization terms is constructed based on the predicted and input values ​​of the preset deep learning model, and a penalty function is constructed based on the weight coefficients between the input and output neurons of the backpropagation network layer. The target loss function is constructed based on the regularization-based loss function and the penalty function. The weight coefficients between the input and output neurons of the backpropagation network layer are updated based on the target loss function until the convergence condition is met. The current weight coefficients between the input and output neurons of the backpropagation network layer are then used as the network parameters of the backpropagation network layer.

6. The energy storage system control method based on a fusion terminal according to claim 5, characterized in that, The target loss function includes: in, Let be the target loss function. and These are the actual value and the predicted value, respectively. Let m be the input feature vector of the x-th sample, and m be the number of samples. Let be the penalty function. The proportionality coefficient of the penalty function, For the loss function based on the regularization term, These are the weight coefficients between the input and output neurons of the backpropagation network layer.

7. A method for regulating an energy storage system based on a converged terminal, applied to a converged terminal, characterized in that, include: Send the environmental parameters and historical load of the target area within the specified time period to the cloud platform; The system receives the day-ahead scheduling plan for the target distribution area generated by the cloud platform based on the environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted load for the target distribution area during different day-ahead scheduling periods and the charging / discharging status of the energy storage units during those periods. Using environmental parameters within the specified time period as input, the predicted power output of the target power distribution area at different times is obtained through the power output prediction model. The real-time power output of the target power distribution area is obtained. Based on the error between the real-time power output of the target power distribution area and the predicted power output at the corresponding time, the energy storage unit of the target power distribution area is controlled to charge and discharge, so that the real-time power output of the target power distribution area matches the corresponding predicted power output. The power output prediction model is obtained by training a preset deep learning model on a cloud platform based on historical environmental parameters and historical power output of different power distribution areas.

8. The energy storage system control method based on a fusion terminal according to claim 7, characterized in that, Based on the error between the real-time output power of the target power distribution area and the predicted output power at the corresponding time, the energy storage units in the target power distribution area are controlled to charge and discharge, including: If the real-time output power of the target power station is lower than the predicted output power of the target power station at the current time, the output power difference between the real-time output power and the predicted output power is determined. The energy storage unit is controlled to discharge based on the output power difference until the real-time output power of the target power station matches the predicted output power at the current sampling time, at which point the energy storage unit stops discharging.

9. The energy storage system control method based on a fusion terminal according to claim 7, characterized in that, The energy storage unit includes at least one battery module, and controlling the energy storage unit in the target area to charge and discharge includes: Obtain the SOH value, which characterizes the health status of each battery module, and the cumulative number of cycles for each battery module; Sort the battery modules by SOH value from high to low, or sort the battery modules by cumulative cycle count from low to high; To determine the target discharge power, during the discharge process of the energy storage unit, target battery modules are determined sequentially from high to low according to the sorting order of each battery module until the total discharge power of each target battery module is not less than the target discharge power.

10. The energy storage system control method based on a fusion terminal according to claim 9, characterized in that, Controlling the energy storage units in the target area to charge and discharge also includes: Energy storage units with SOH values ​​below a preset threshold are identified as target energy storage units. During the charging process of the energy storage unit, the charging current of the target energy storage unit is controlled to be no greater than a set charging current, which is less than the rated charging current of the energy storage unit.

11. A control device for an energy storage system based on a fusion terminal, employing the control method for an energy storage system based on a fusion terminal as described in any one of claims 1-6, characterized in that, The device comprises: The load forecasting module is configured to obtain the environmental parameters and historical load of the target transformer area within a specified time period through the target fusion terminal, and obtain the predicted load of the target transformer area by using the environmental parameters and the historical load as input and a pre-trained load forecasting model. The plan generation module is configured to acquire time-of-use electricity price data, generate a day-ahead scheduling plan for the target distribution area based on the predicted load and the time-of-use electricity price data, and send the day-ahead scheduling plan to the target fusion terminal; wherein, the load prediction model is obtained by training a preset deep learning model based on different historical loads of different distribution areas under different environmental parameters, using an improved parameter optimization algorithm. The improved parameter optimization algorithm is used to update the model parameters of the deep learning model during the training process by calculating the first moment estimate and second moment estimate of the iterative gradient.

12. A control device for an energy storage system based on a converged terminal, employing the control method for an energy storage system based on a converged terminal as described in any one of claims 7-10, characterized in that, The device comprises: The data transmission module is configured to send the environmental parameters and historical load of the target area within a specified time period to the cloud platform; The data receiving module is configured to receive the day-ahead scheduling plan of the target distribution area generated by the cloud platform based on the environmental parameters and historical load. The day-ahead scheduling plan includes at least the predicted load of the target distribution area in different day-ahead scheduling periods and the charging and discharging status of the energy storage unit in different day-ahead scheduling periods. The control module is configured to take environmental parameters within the specified time period as input, obtain the predicted power output of the target power distribution area at different times through a power output prediction model, acquire the real-time power output of the target power distribution area, and control the charging and discharging of the energy storage unit of the target power distribution area based on the error between the real-time power output of the target power distribution area and the predicted power output at the corresponding time, so that the real-time power output of the target power distribution area matches the corresponding predicted power output; wherein, the power output prediction model is obtained by the cloud platform after training a preset deep learning model based on the historical environmental parameters and historical power output of different power distribution areas.

13. A machine-readable storage medium storing instructions thereon, characterized in that, When executed by a processor, the instruction causes the processor to be configured to perform the energy storage system control method based on a fusion terminal as described in any one of claims 1-6, or to perform the energy storage system control method based on a fusion terminal as described in claims 7-10.

14. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the energy storage system control method based on a fusion terminal as described in any one of claims 1-6, or implements the steps of the energy storage system control method based on a fusion terminal as described in claims 7-10.

Images