Cooperative control method of multi-mode energy storage system

Abstract

The invention discloses a cooperative control method for a multi-modal energy storage system, and relates to the technical field of energy storage control, and the method comprises the steps: S1, carrying out the data collection, obtaining multi-modal initial data and a real-time data set, and carrying out the preprocessing of the multi-modal initial data, and obtaining synchronous multi-modal data; s2, performing preliminary feature extraction and embedded mapping on the synchronous multi-modal data, and performing feature fusion and vector batch normalization processing through an InTense fusion network to obtain a final fusion feature vector; s3, contribution degree calculation is carried out based on the final fusion feature vector, and a modal contribution proportion set is obtained; s4, establishing a linear regression model, and calculating through a weighted least square method to obtain predicted load data; s5, in combination with the modal contribution proportion set, performing optimization calculation on the mixed integer quadratic programming model through a particle swarm optimization algorithm to obtain an energy storage charging and discharging strategy; and S6, converting the energy storage charging and discharging strategy to obtain regulation and control instruction data, and completing real-time control.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage control, and in particular to a collaborative control method for a multi-modal energy storage system. Background Art

[0002] In recent years, thanks to innovative breakthroughs in the fields of large-scale application of renewable energy, upgrading of smart grid technology, and continuous reduction in energy storage costs, energy storage applications have also evolved from single energy storage to multi-type integrated energy systems. Large public buildings such as commercial complexes, data centers, hospital buildings, and transportation hubs are characterized by high energy consumption and coordinated demand for electric and thermal loads, which places higher demands on the flexibility and economy of energy storage systems. Traditional single-form energy storage solutions are difficult to adapt to the complex energy flow interaction scenarios of buildings. Multimodal energy storage systems can effectively improve energy utilization efficiency and reduce system operating costs by integrating multiple energy storage resources of electricity and heat. Therefore, the study of analytical control methods for multimodal energy storage systems for large public buildings has application prospects.

[0003] Currently, a Chinese invention patent application with application number CN202410513894.8 discloses a method and system for energy storage safety early warning based on multimodal sensing. The application includes: obtaining a historical monitoring data set of an energy storage battery pack, including multiple monitoring indicators; establishing early warning association levels, including first-level early warning association, second-level early warning association, and third-level early warning association; deploying a first-level detection device to obtain a first-level detection data set; deploying a second-level detection device to obtain a second-level detection data set; deploying a third-level detection device to obtain a third-level detection data set; performing anomaly identification based on an anomaly recognition model to generate an anomaly identification result with an anomaly level identifier; sending the result to a management terminal to issue a safety early warning based on the anomaly level. This application solves the technical problem that existing technologies generally focus only on the monitoring and analysis of single-modal data, lack comprehensive processing of multiple monitoring indicators, and lack a classification and early warning mechanism for different abnormal situations, resulting in poor early warning accuracy. Although the application analyzes from a multimodal perspective, it only focuses on abnormal situation identification and lacks analysis of normal operating conditions and control schemes. Summary of the Invention

[0004] The technical problem solved by the present invention is that most existing energy storage technologies usually focus on the analysis and control of a single energy type and lack multi-modal analysis. The existing technologies that use multi-modal analysis are mostly aimed at fault identification rather than normal operation regulation. At the same time, the existing technologies fail to reasonably solve the time scale differences and physical property coupling problems between high-frequency transient power data and low-frequency slowly varying thermal data, resulting in poor cross-modal data fusion effects. The ability to differentially capture the trend characteristics of power data and thermal data is insufficient, and there is a lack of specific quantitative methods to calculate the contribution ratio of different modes in optimization decisions, resulting in insufficient adaptability and real-time performance of the output final energy storage control strategy.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: A coordinated control method for a multi-modal energy storage system, comprising: Step S1, data acquisition is performed to obtain multimodal initial data and real-time data sets, and preprocessing is performed to obtain synchronized multimodal data; Step S2: perform preliminary feature extraction and embedding mapping on the synchronized multimodal data, perform feature fusion and vector-level batch normalization processing through the InTense fusion network to obtain the final fused feature vector; Step S3, calculating the contribution based on the final fused feature vector to obtain a modal contribution ratio set; Step S4, establishing a linear regression model and obtaining predicted load data by weighted least squares method; Step S5: combining the modal contribution ratio set, optimizing the mixed integer quadratic programming model using a particle swarm optimization algorithm to obtain an energy storage charge and discharge strategy, which includes battery charge and discharge power and SOC adjustment targets; In step S6, the energy storage charge and discharge strategy is converted into control instruction data, and the battery management system and the energy storage converter adjust the charge and discharge current based on the control instruction data to complete real-time control.

[0006] Preferably, in step S1, the multimodal initial data includes power time series data and thermal time series data; The power time series data includes voltage data, current data, frequency data, lithium battery SOC data and capacitor charge and discharge power data, and the processing logic includes: The voltage data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a voltage transformer; The current data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a current transformer; Frequency data is collected at the grid interface common connection point (PCC) through a frequency counter; The battery management system (BMS) collects the SOC data of the lithium battery from the energy storage battery pack; The power sensor is used to collect the charge and discharge power data of the supercapacitor group.

[0007] Preferably, the thermal time series data includes the liquid level of the cold storage tank, the cold storage temperature difference, the indoor temperature and the indoor humidity, and the processing logic includes: The liquid level height of the cold storage tank is obtained by collecting the data inside the cold storage tank through the liquid level sensor; The supply and return water temperatures are collected by temperature sensors deployed on the chilled water supply and return pipes, and the difference between the return and supply water temperatures is calculated to obtain the cold storage temperature difference. The indoor temperature and humidity are collected by temperature and humidity sensors deployed in the interior area of ​​the building; The real-time data set includes the total active load value, the mass flow rate of the cold storage liquid, the real-time electricity price and the real-time weather temperature. The processing logic includes: The total active load value is obtained by collecting data from the grid interface common connection point (PCC) and the AC side output terminal of the energy storage system main converter (PCS) through power sensors; The mass flow rate of the cold storage liquid is obtained by collecting the cold storage tank inlet and outlet pipes through an electromagnetic flowmeter.

[0008] Preferably, the multimodal initial data is preprocessed to obtain synchronized multimodal data, wherein the synchronized multimodal data includes power synchronized data and thermal synchronized data, and the processing logic includes: Perform timestamp alignment on the power time series data and thermal time series data in the multimodal initial data to obtain aligned time series data, and then perform resampling on the aligned time series data using linear interpolation to obtain the same-frequency time series data. Based on the 3σ principle, outliers are replaced on the same-frequency time series data to obtain synchronized multimodal data.

[0009] Preferably, a dual-channel 1DCNN network is used to perform preliminary feature extraction on the synchronous multimodal data, and embedding mapping processing is performed to obtain the power modal feature embedding vector and the thermal modal feature embedding vector. The processing logic includes: A sliding window convolution is performed on the power synchronization data and the thermal synchronization data respectively through a shared 1D convolution layer to obtain a power local feature map and a thermal local feature map. The shared 1D convolution layer includes 64 1D convolution kernels with a size of 3×1 and a stride of 1; The power branch convolution layer processes the power local feature map to obtain the power modal feature embedding vector. The processing logic includes: The high-frequency fluctuation features are obtained by performing a second convolution calculation on the local power feature map using 32 1D convolution kernels with a size of 5×1 and a stride of 1. The high-frequency fluctuation features are nonlinearly transformed using the ReLU activation function. Global average pooling is used for the first compression process to obtain 32-dimensional power dimensionality reduction features. The power dimensionality reduction features are mapped through the first fully connected layer to obtain the power modal feature embedding vector. The thermal branch convolution layer processes the thermal local feature map to obtain the thermal modal feature embedding vector. The processing logic includes: The second convolution calculation is performed on the thermal local feature map through 32 1D convolution kernels with a size of 7×1 and a stride of 1 to obtain the low-frequency fluctuation feature. The low-frequency fluctuation feature is nonlinearly transformed through the LeakyReLU activation function. The second compression processing is performed using global average pooling to obtain a 32-dimensional thermal dimensionality reduction feature. The thermal dimensionality reduction feature is mapped through the second fully connected layer mapping to obtain the thermal modal feature embedding vector.

[0010] Preferably, in step S2, feature fusion and vector-level batch normalization are performed on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain a final fused feature vector. The processing logic includes: Perform vector outer product operation on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain the modal interaction tensor, and perform Z-score normalization and vector splicing to obtain the combined input vector, which is expressed as follows: Combined input vector = [electrical modal feature embedding vector, thermal modal feature embedding vector, modal interaction tensor]; The InTense fusion network is used to perform feature fusion and batch normalization on the combined input vector. The processing logic includes: The combined input vector is multiplied by the first weight matrix in the linear combination layer, added to the first bias term, and then nonlinearly transformed by the ReLU function to obtain the global fusion feature. The dimension of the first weight matrix is ​​256×1088. The modal interaction tensor is multiplied by the second weight matrix in the tensor product interaction layer, added to the second bias term, and then nonlinearly transformed by the ReLU function to obtain the interaction reinforcement feature. The dimension of the second weight matrix is ​​256×1024. The global fusion features and interactive enhancement features are added element by element to obtain the preliminary fusion features, and the preliminary fusion features are batch normalized by the vector-level batch normalization (VBN) algorithm to obtain the final fusion feature vector.

[0011] Preferably, in step S3, the contribution degree is calculated based on the final fused feature vector to obtain a modal contribution ratio set, and the processing logic includes: The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the power modal feature embedding vector to obtain a first sub-matrix, where the dimension of the first sub-matrix is ​​256×32, and the 32 dimensions corresponding to the power modal feature embedding vector are the first 32 columns of the combined input vector; The square root of the sum of squares of the elements in each row of the first submatrix is ​​calculated to obtain the first L2 norm set, and the arithmetic mean of the corresponding norms of each row in the first L2 norm set is taken to obtain the original value of the power modal contribution; The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the thermal modal feature embedding vector to obtain a second sub-matrix, where the dimension of the second sub-matrix is ​​256×32, and the 32 dimensions corresponding to the thermal modal feature embedding vector are the 33rd to 64th columns of the combined input vector; The square root of the sum of squares of the elements in each row of the second submatrix is ​​calculated to obtain the second L2 norm set, and the arithmetic mean of the corresponding norms of each row in the second L2 norm set is taken to obtain the original value of the thermal modal contribution; The square root of the sum of squares of the elements in each row of the second weight matrix is ​​calculated to obtain the third L2 norm set, and the arithmetic mean of the corresponding norms of each row in the third L2 norm set is taken to obtain the original value of the modal interaction contribution; A modal contribution ratio set is obtained by summing and normalizing the original values ​​of the power modal contribution, the thermal modal contribution, and the modal interaction contribution. The modal contribution ratio set includes the power modal contribution ratio, the thermal modal contribution ratio, and the modal interaction contribution ratio. The calculation logic of the summation and normalization includes: Normalized base value = original value of power modal contribution + original value of thermal modal contribution + original value of modal interaction contribution; Power modal contribution ratio = power modal contribution original value / normalized base value; Thermal mode contribution ratio = thermal mode contribution original value / normalized base value; Modal interaction contribution ratio = original value of modal interaction contribution / normalized base value.

[0012] Preferably, in step S4, a linear regression model is established to obtain the predicted load data by weighted least squares method. The processing logic includes: The final fusion feature vector, external features and historical loads are Z-score normalized and vector concatenated to obtain linear regression input items and establish a linear regression model; The external features include real-time electricity prices and real-time weather temperatures; The historical load includes historical power load data and historical thermal load data within 14 days, divided into 5-minute steps. The historical power load data includes historical total active load values, and the historical thermal load data includes historical cooling load values. The calculation expression of the linear regression model is: ; Where t represents time, P(t) represents the predicted load data, represents the linear regression input, represents the final fusion feature vector, D(t) represents the external feature, H(t) represents the historical load, represents the linear regression weight matrix, represents the linear regression bias term; The linear regression model is fitted and calculated by weighted least squares (WLS) to obtain the predicted load data for the next 6 hours. The predicted load data includes predicted power load data and predicted thermal load data, which are divided into 72 time steps with a step length of 5 minutes.

[0013] Preferably, in step S5, a particle swarm optimization calculation is performed based on the modal contribution ratio set and the predicted load data to obtain an energy storage charging and discharging strategy. The processing logic includes: With the goal of minimizing the electricity purchase cost, a penalty term is constructed based on the power mode contribution ratio and the thermal mode contribution ratio, and a mixed integer quadratic programming (MIQP) model is established. Its calculation expression is: ; in, represents the minimum electricity purchase cost, U(t) represents the electricity purchase price, represents the purchased power, λ represents the penalty coefficient, represents the power mode contribution ratio, represents the power forecast error, represents the contribution ratio of thermal mode, represents the thermal prediction error; The mixed integer quadratic programming model is optimized by combining the APSO particle swarm optimization algorithm with the modal interaction contribution ratio. The processing logic includes: The inertia weight is calculated by linear mapping based on the modal interaction contribution ratio. The speed update formula of the particle swarm optimization algorithm is adjusted by the inertia weight to obtain the particle swarm search parameter. The energy storage charging and discharging strategy is obtained by iterative calculation based on the particle swarm search parameter. The energy storage charging and discharging strategy includes the battery charging and discharging power and SOC adjustment target, divided into 5-minute steps, with a total of 72 time steps; When the battery charge and discharge power is a positive number, it indicates charging, and when it is a negative number, it indicates discharging; The calculation expression of the inertia weight is: ; in, represents the inertia weight, represents the initial global search weight, represents the local convergence weight, Indicates the contribution ratio of modal interaction.

[0014] Preferably, in step S6, the energy storage charge and discharge strategy is converted into JSON format to obtain control instruction data, the control instruction data is sent to the energy storage system via the HTTP protocol, and the control instruction data is transmitted to the battery management system (BMS) and the energy storage converter (PCS) at the energy storage system device end via the CAN protocol. The battery management system and the energy storage converter adjust the charge and discharge current based on the control instruction data to complete real-time control.

[0015] The beneficial effects of the present invention: The heterogeneity problem of high-frequency transient data of electricity and low-frequency slowly varying data of heat is solved by frequency preprocessing. Taking into full consideration that the physical properties of electricity data and thermal data after frequency preprocessing are still different, the present invention adopts a 1DCNN network with differentiated convolutional layers to strengthen the characteristics of electricity and heat respectively. The weights of electricity, heat and interaction factors in load forecasting are accurately quantified by calculating the modal contribution ratio set, avoiding fixed thresholds that rely on experience. The modal contribution ratio set is combined with a mature algorithm, and a multi-objective balance of minimizing electricity purchase costs, ensuring the stability of grid frequency regulation and maintaining the safety range of energy storage SOC is achieved through mixed integer quadratic programming and particle swarm optimization algorithm. It adapts to the dynamic changes of energy loads in large public buildings and meets the needs of real-time control and reducing economic costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the basic flow of a collaborative control method for a multi-modal energy storage system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0018] Reference Figure 1 , as an embodiment of the present invention, provides a coordinated control method for a multi-modal energy storage system, comprising: Step S1, data acquisition is performed to obtain multimodal initial data and real-time data sets, and preprocessing is performed to obtain synchronized multimodal data; Step S2: perform preliminary feature extraction and embedding mapping on the synchronized multimodal data, perform feature fusion and vector-level batch normalization processing through the InTense fusion network to obtain the final fused feature vector; Step S3, calculating the contribution based on the final fused feature vector to obtain a modal contribution ratio set; Step S4, establishing a linear regression model and obtaining predicted load data by weighted least squares method; Step S5: combining the modal contribution ratio set, optimizing the mixed integer quadratic programming model using a particle swarm optimization algorithm to obtain an energy storage charge and discharge strategy, which includes battery charge and discharge power and SOC adjustment targets; In step S6, the energy storage charge and discharge strategy is converted into control instruction data, and the battery management system and the energy storage converter adjust the charge and discharge current based on the control instruction data to complete real-time control.

[0019] In this embodiment, in step S1, the multimodal initial data includes power time series data and thermal time series data; The power time series data includes voltage data, current data, frequency data, lithium battery SOC data and capacitor charge and discharge power data, and the processing logic includes: The voltage data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a voltage transformer; The current data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a current transformer; Frequency data is collected at the grid interface common connection point (PCC) through a frequency counter; The battery management system (BMS) collects the SOC data of the lithium battery from the energy storage battery pack; The power sensor is used to collect the charge and discharge power data of the supercapacitor group.

[0020] The national standard GB / T15945-2008 stipulates that during normal operation, the frequency of a building's power grid must deviate from the nominal value of 50Hz by ±0.2Hz. Therefore, in addition to voltage and current, this application selects real-time acquisition of building power grid frequency data as a characteristic variable, which facilitates frequency regulation control of the energy storage system and grid stability maintenance.

[0021] In this embodiment, the thermal time series data includes the liquid level of the cold storage tank, the cold storage temperature difference, the indoor temperature and the indoor humidity. The processing logic includes: The liquid level height of the cold storage tank is obtained by collecting the data inside the cold storage tank through the liquid level sensor; The supply and return water temperatures are collected by temperature sensors deployed on the chilled water supply and return pipes, and the difference between the return and supply water temperatures is calculated to obtain the cold storage temperature difference. The indoor temperature and humidity are collected by temperature and humidity sensors deployed in the interior area of ​​the building; The real-time data set includes the total active load value, the mass flow rate of the cold storage liquid, the real-time electricity price and the real-time weather temperature. The processing logic includes: The total active load value is obtained by collecting data from the grid interface common connection point (PCC) and the AC side output terminal of the energy storage system main converter (PCS) through power sensors; The mass flow rate of the cold storage liquid is obtained by collecting the cold storage tank inlet and outlet pipes through an electromagnetic flowmeter.

[0022] Among them, the cold storage temperature difference = return water temperature - supply water temperature, and the real-time electricity price and real-time weather temperature are obtained from the power grid data service platform and the weather data service platform respectively.

[0023] In this embodiment, the multimodal initial data is preprocessed to obtain synchronized multimodal data, which includes power synchronized data and thermal synchronized data. The processing logic includes: Perform timestamp alignment on the power time series data and thermal time series data in the multimodal initial data to obtain aligned time series data, and then perform resampling on the aligned time series data using linear interpolation to obtain the same-frequency time series data. Based on the 3σ principle, outliers are replaced on the same-frequency time series data to obtain synchronized multimodal data.

[0024] Among them, the preprocessed synchronized multimodal data includes timestamp alignment and removes significant outliers in the real-time electrical and thermal acquisition values, which is conducive to reducing error infection interference.

[0025] In this embodiment, a dual-channel 1DCNN network is used to perform preliminary feature extraction on the synchronous multimodal data, and embedding mapping processing is performed to obtain the power modal feature embedding vector and the thermal modal feature embedding vector. The processing logic includes: A sliding window convolution is performed on the power synchronization data and the thermal synchronization data respectively through a shared 1D convolution layer to obtain a power local feature map and a thermal local feature map. The shared 1D convolution layer includes 64 1D convolution kernels with a size of 3×1 and a stride of 1; The power branch convolution layer processes the power local feature map to obtain the power modal feature embedding vector. The processing logic includes: The high-frequency fluctuation features are obtained by performing a second convolution calculation on the local power feature map using 32 1D convolution kernels with a size of 5×1 and a stride of 1. The high-frequency fluctuation features are nonlinearly transformed using the ReLU activation function. Global average pooling is used for the first compression process to obtain 32-dimensional power dimensionality reduction features. The power dimensionality reduction features are mapped through the first fully connected layer to obtain the power modal feature embedding vector. The thermal branch convolution layer processes the thermal local feature map to obtain the thermal modal feature embedding vector. The processing logic includes: The second convolution calculation is performed on the thermal local feature map through 32 1D convolution kernels with a size of 7×1 and a stride of 1 to obtain the low-frequency fluctuation feature. The low-frequency fluctuation feature is nonlinearly transformed through the LeakyReLU activation function. The second compression processing is performed using global average pooling to obtain a 32-dimensional thermal dimensionality reduction feature. The thermal dimensionality reduction feature is mapped through the second fully connected layer mapping to obtain the thermal modal feature embedding vector.

[0026] Among them, the frequency synchronization processing in the aforementioned preprocessing step only solves the problems of timestamp alignment and sampling density consistency, and obtains power synchronization data and thermal synchronization data with timestamp alignment and sampling density consistency. However, the power data and thermal data of the energy storage system still have differences in physical characteristics such as fluctuation frequency, amplitude range and noise distribution. Even if the sampling is performed at the same frequency, the power data may show transient fluctuations such as millisecond-level spikes due to sudden changes in the load, while the thermal data usually shows a low-frequency slow-changing trend with a gradual change of minutes. Therefore, this application uses a 1DCNN network with differentiated convolutional layers for preliminary feature extraction, which strengthens the high-frequency fluctuation characteristics of electricity and the low-frequency slow-changing characteristics of heat respectively. At the same time, the lightweight architecture does not introduce complex mechanisms such as attention commonly used in existing 1DCNN networks. It only focuses on preliminary feature extraction, and the in-depth feature analysis and processing are handed over to the subsequent Intense fusion network. The reasonable division of labor is conducive to improving computing efficiency and real-time performance.

[0027] In this embodiment, step S2 performs feature fusion and vector-level batch normalization on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain a final fused feature vector. The processing logic includes: Perform vector outer product operation on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain the modal interaction tensor, and perform Z-score normalization and vector splicing to obtain the combined input vector, which is expressed as follows: Combined input vector = [electrical modal feature embedding vector, thermal modal feature embedding vector, modal interaction tensor]; The InTense fusion network is used to perform feature fusion and batch normalization on the combined input vector. The processing logic includes: The combined input vector is multiplied by the first weight matrix in the linear combination layer, added to the first bias term, and then nonlinearly transformed by the ReLU function to obtain the global fusion feature. The dimension of the first weight matrix is ​​256×1088. The modal interaction tensor is multiplied by the second weight matrix in the tensor product interaction layer, added to the second bias term, and then nonlinearly transformed by the ReLU function to obtain the interaction reinforcement feature. The dimension of the second weight matrix is ​​256×1024. The global fusion features and interactive enhancement features are added element by element to obtain the preliminary fusion features, and the preliminary fusion features are batch normalized by the vector-level batch normalization (VBN) algorithm to obtain the final fusion feature vector.

[0028] Among them, the calculation expression for the final fusion feature vector obtained by vector-level batch normalization is: ; in, Represents the final fusion feature vector, Moving average, represents the moving variance, Indicates the minimum value, take 1 10^(-5); The vector-level batch normalization algorithm needs to call the moving average and moving variance stored in the training phase during calculation. The moving average and moving variance are calculated by updating the initial value of the moving average and the initial value of the moving variance. In this embodiment, the initial value of the moving average is set to 0, the initial value of the moving variance is set to 1, the training decay coefficient is set to 0.99, and the moving variance is calculated based on the updated moving average. The update rule expression of the moving average is: ; in, represents the exponential moving average (EMA) decay coefficient, Represents the mean value of the feature dimension of the current batch, Represents the moving average value obtained by the current update, Represents the moving average obtained from the previous update.

[0029] Among them, the existing technology requires linear scaling after batch normalization processing. This application adopts the vector-level batch normalization (VBN) algorithm. It is precisely because the advantage of the VBN algorithm is that it removes scaling and offset, which is conducive to the interaction of power / thermal features of different modes at a unified scale, retaining the original feature distribution after vector-level normalization, and reducing high-order fusion deviation.

[0030] In this embodiment, step S3 calculates the contribution based on the final fused feature vector to obtain a modal contribution ratio set. The processing logic includes: The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the power modal feature embedding vector to obtain a first sub-matrix, where the dimension of the first sub-matrix is ​​256×32, and the 32 dimensions corresponding to the power modal feature embedding vector are the first 32 columns of the combined input vector; The square root of the sum of squares of the elements in each row of the first submatrix is ​​calculated to obtain the first L2 norm set, and the arithmetic mean of the corresponding norms of each row in the first L2 norm set is taken to obtain the original value of the power modal contribution; The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the thermal modal feature embedding vector to obtain a second sub-matrix, where the dimension of the second sub-matrix is ​​256×32, and the 32 dimensions corresponding to the thermal modal feature embedding vector are the 33rd to 64th columns of the combined input vector; The square root of the sum of squares of the elements in each row of the second submatrix is ​​calculated to obtain the second L2 norm set, and the arithmetic mean of the corresponding norms of each row in the second L2 norm set is taken to obtain the original value of the thermal modal contribution; The square root of the sum of squares of the elements in each row of the second weight matrix is ​​calculated to obtain the third L2 norm set, and the arithmetic mean of the corresponding norms of each row in the third L2 norm set is taken to obtain the original value of the modal interaction contribution; A modal contribution ratio set is obtained by summing and normalizing the original values ​​of the power modal contribution, the thermal modal contribution, and the modal interaction contribution. The modal contribution ratio set includes the power modal contribution ratio, the thermal modal contribution ratio, and the modal interaction contribution ratio. The calculation logic of the summation and normalization includes: Normalized base value = original value of power modal contribution + original value of thermal modal contribution + original value of modal interaction contribution; Power modal contribution ratio = power modal contribution original value / normalized base value; Thermal mode contribution ratio = thermal mode contribution original value / normalized base value; Modal interaction contribution ratio = original value of modal interaction contribution / normalized base value.

[0031] In this embodiment, step S4 is to establish a linear regression model and calculate the predicted load data using the weighted least squares method. The processing logic includes: The final fusion feature vector, external features and historical loads are Z-score normalized and vector concatenated to obtain linear regression input items and establish a linear regression model; The external features include real-time electricity prices and real-time weather temperatures; The historical load includes historical power load data and historical thermal load data within 14 days, divided into 5-minute steps. The historical power load data includes historical total active load values, and the historical thermal load data includes historical cooling load values. The calculation expression of the linear regression model is: ; Where t represents time, P(t) represents the predicted load data, represents the linear regression input, represents the final fusion feature vector, D(t) represents the external feature, H(t) represents the historical load, represents the linear regression weight matrix, represents the linear regression bias term; Among them, when H(t) is used as the input of the historical load linear regression model, the power and heat load values ​​at time t in the past 14 days are extracted with a 24-hour cycle.

[0032] The linear regression model is fitted and calculated by weighted least squares (WLS) to obtain the predicted load data for the next 6 hours. The predicted load data includes predicted power load data and predicted thermal load data, which are divided into 72 time steps with a step length of 5 minutes.

[0033] The logic for fitting the linear regression model using the weighted least squares (WLS) method includes: The objective vector of the weighted least squares method is expanded into a two-dimensional objective matrix Y containing power load and thermal load, which is expressed as follows: Y=[forecasted power load data, forecasted thermal load data] The WLS algorithm is used to minimize the weighted residual square sum of fitting calculations, and the regression coefficients are solved to obtain the linear regression weight matrix and linear regression bias term corresponding to the predicted power load and predicted thermal load. The predicted power load data and predicted thermal load data are obtained by substituting the linear regression model back. Each column in the linear regression weight matrix corresponds to the coefficient of the predicted electric load and the predicted thermal load; The linear regression bias term is in vector form, and each element corresponds to the bias between the predicted electric load and the predicted thermal load.

[0034] The relationship between real-time power load data and historical power load data differs only in temporal state and can be retrieved from the real-time data set in step S1. The relationship between real-time thermal load data and historical thermal load data also differs only in temporal state and cannot be retrieved directly but must be calculated. However, the calculation method is the same, and the calculation expression for thermal load data is: ; in, represents the thermal load data, c is the specific heat capacity of the cold storage liquid, m(t) is the mass flow rate of the cold storage liquid, Indicates the cold storage temperature difference.

[0035] Forecasted power load data differs from real-time / historical power load data not only in temporal form but also in that it is derived from weighted least squares fitting estimates rather than from measured data. Similarly, forecasted thermal load data is derived from weighted least squares fitting estimates, a calculation method different from non-real-time / historical thermal load data.

[0036] In this embodiment, step S5 combines the modal contribution ratio set and the predicted load data to perform particle swarm optimization calculation to obtain the energy storage charging and discharging strategy. The processing logic includes: With the goal of minimizing the electricity purchase cost, a penalty term is constructed based on the power mode contribution ratio and the thermal mode contribution ratio, and a mixed integer quadratic programming (MIQP) model is established. Its calculation expression is: ; in, represents the minimum electricity purchase cost, U(t) represents the electricity purchase price, represents the purchased power, λ represents the penalty coefficient, and the dimension of λ is yuan / (kW·h). It represents the penalty cost corresponding to the unit power prediction error. represents the power mode contribution ratio, represents the power forecast error, represents the contribution ratio of thermal mode, represents the thermal prediction error; Among them, the power / heat prediction error ( ) is not a fixed value item obtained by offline training in the traditional MIQP model, but is obtained by subtracting the predicted power / heat load data of the previous cycle (that is, the prediction for the current time 5 minutes ago, historical prediction data) from the real-time power / heat load data at the current time. This application performs a rolling optimization every 5 minutes (the step size of the corresponding historical load data is also 5 minutes) to achieve real-time data feedback.

[0037] Among them, the constraints of MIQP include energy balance, energy storage SOC range, energy storage rated capacity and maximum power purchase; The calculation expression of energy balance is: ; in, Indicates the total active load value, Indicates the purchased power, Indicates battery charge and discharge power; Energy storage SOC range is 20% to 80%; Among them, the thermal load consumes electricity through thermal equipment such as refrigeration equipment, and the thermal demand is unified into electricity consumption, which is indirectly reflected in the total power load.

[0038] The mixed integer quadratic programming model is optimized by combining the APSO particle swarm optimization algorithm with the modal interaction contribution ratio. The processing logic includes: The inertia weight is calculated by linear mapping based on the modal interaction contribution ratio. The speed update formula of the particle swarm optimization algorithm is adjusted by the inertia weight to obtain the particle swarm search parameter. The energy storage charging and discharging strategy is obtained by iterative calculation based on the particle swarm search parameter. The energy storage charging and discharging strategy includes the battery charging and discharging power and SOC adjustment target, divided into 5-minute steps, with a total of 72 time steps; When the battery charge and discharge power is a positive number, it indicates charging, and when it is a negative number, it indicates discharging; The calculation expression of the inertia weight is: ; in, represents the inertia weight, represents the initial global search weight, represents the local convergence weight, Indicates the contribution ratio of modal interaction.

[0039] The initial population size of the particle swarm is set to 50, the maximum number of iterations is set to 120, and the convergence threshold is set to 1. 10^(-3); Among them, this application uses the modal contribution ratio set as an important weight parameter, constructs the penalty term of the mixed integer quadratic programming (MIQP) model with the power modal contribution ratio and the thermal modal contribution ratio, and dynamically adjusts the inertia weight of the particle swarm optimization (APSO) through the modal interaction contribution ratio. This helps to solve the problem that traditional fixed-parameter particle swarms are prone to falling into local optimality, and improves the accuracy of the balance between the economy and reliability of the strategy. The least squares method, mixed integer quadratic programming model and adaptive particle swarm optimization algorithm used at the same time are all mature algorithms. This application calculates the final fused eigenvector and modal contribution ratio set through the aforementioned steps. The final fused eigenvector and modal contribution ratio set that reflect the physical modal characteristics are cleverly combined with the optimization algorithm parameters in depth, achieving the goal of simplifying the computational complexity while maintaining accuracy, which is conducive to improving the adaptability and robustness of this application and obtaining an energy storage charging and discharging strategy that takes into account both physical constraints and economic goals.

[0040] In this embodiment, step S6 converts the energy storage charge and discharge strategy into JSON format to obtain control instruction data, sends the control instruction data to the energy storage system via the HTTP protocol, and transmits the control instruction data to the battery management system (BMS) and power storage converter (PCS) on the energy storage system device side via the CAN protocol. The BMS and power storage converter adjust the charge and discharge current based on the control instruction data to achieve real-time control.

[0041] Those skilled in the art will appreciate that embodiments of the present invention may provide methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0042] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A collaborative control method for a multi-modal energy storage system, characterized in that: include: Step S1: data acquisition is performed to obtain multimodal initial data and a real-time data set, and the multimodal initial data is preprocessed to obtain synchronized multimodal data; Step S2: perform preliminary feature extraction and embedding mapping on the synchronized multimodal data, perform feature fusion and vector-level batch normalization processing through the InTense fusion network to obtain the final fused feature vector; Step S3, calculating the contribution based on the final fused feature vector to obtain a modal contribution ratio set; Step S4, establishing a linear regression model and obtaining predicted load data by weighted least squares method; Step S5: combining the modal contribution ratio set, optimizing the mixed integer quadratic programming model using a particle swarm optimization algorithm to obtain an energy storage charge and discharge strategy, which includes battery charge and discharge power and SOC adjustment targets; In step S6, the energy storage charge and discharge strategy is converted into control instruction data, and the battery management system and the energy storage converter adjust the charge and discharge current based on the control instruction data to complete real-time control.

2. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: Step S1, the multimodal initial data includes power time series data and thermal time series data; The power time series data includes voltage data, current data, frequency data, lithium battery SOC data and capacitor charge and discharge power data, and the processing logic includes: The voltage data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a voltage transformer; The current data is collected from the AC output terminal of the main current converter (PCS) of the energy storage system through a current transformer; Frequency data is collected at the grid interface common connection point (PCC) through a frequency counter; The battery management system (BMS) collects the SOC data of the lithium battery from the energy storage battery pack; The power sensor is used to collect the charge and discharge power data of the supercapacitor group.

3. The coordinated control method of a multi-modal energy storage system according to claim 2, characterized in that: The thermal time series data includes the liquid level of the cold storage tank, the cold storage temperature difference, the indoor temperature and the indoor humidity. The processing logic includes: The liquid level height of the cold storage tank is obtained by collecting the data inside the cold storage tank through the liquid level sensor; The supply and return water temperatures are collected by temperature sensors deployed on the chilled water supply and return pipes, and the cold storage temperature difference is calculated by taking the difference between the return and supply temperatures. The indoor temperature and humidity are collected by temperature and humidity sensors deployed in the interior area of ​​the building; The real-time data set includes the total active load value, the mass flow rate of the cold storage liquid, the real-time electricity price and the real-time weather temperature. The processing logic includes: The total active load value is obtained by collecting data from the grid interface common connection point (PCC) and the AC side output terminal of the energy storage system main converter (PCS) through power sensors; The mass flow rate of the cold storage liquid is obtained by collecting the cold storage tank inlet and outlet pipes through an electromagnetic flowmeter.

4. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: Preprocessing the multimodal initial data to obtain synchronized multimodal data, wherein the synchronized multimodal data includes power synchronized data and thermal synchronized data. The processing logic includes: Perform timestamp alignment on the power time series data and thermal time series data in the multimodal initial data to obtain aligned time series data, and then perform resampling on the aligned time series data using linear interpolation to obtain the same-frequency time series data. Based on the 3σ principle, outliers are replaced on the same-frequency time series data to obtain synchronized multimodal data.

5. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: The dual-channel 1DCNN network is used to perform preliminary feature extraction on the synchronous multimodal data, and embedding mapping processing is performed to obtain the power modal feature embedding vector and the thermal modal feature embedding vector. The processing logic includes: A sliding window convolution is performed on the power synchronization data and the thermal synchronization data respectively through a shared 1D convolution layer to obtain a power local feature map and a thermal local feature map. The shared 1D convolution layer includes 64 1D convolution kernels with a size of 3×1 and a stride of 1; The power branch convolution layer processes the power local feature map to obtain the power modal feature embedding vector. The processing logic includes: The high-frequency fluctuation features are obtained by performing a second convolution calculation on the local power feature map using 32 1D convolution kernels with a size of 5×1 and a stride of 1. The high-frequency fluctuation features are nonlinearly transformed using the ReLU activation function. Global average pooling is used for the first compression process to obtain 32-dimensional power dimensionality reduction features. The power dimensionality reduction features are mapped through the first fully connected layer to obtain the power modal feature embedding vector. The thermal branch convolution layer processes the thermal local feature map to obtain the thermal modal feature embedding vector. The processing logic includes: The second convolution calculation is performed on the thermal local feature map through 32 1D convolution kernels with a size of 7×1 and a stride of 1 to obtain the low-frequency fluctuation feature. The low-frequency fluctuation feature is nonlinearly transformed through the LeakyReLU activation function. The second compression processing is performed using global average pooling to obtain a 32-dimensional thermal dimensionality reduction feature. The thermal dimensionality reduction feature is mapped through the second fully connected layer mapping to obtain the thermal modal feature embedding vector.

6. The coordinated control method of a multi-modal energy storage system according to claim 1, characterized in that: Step S2: Perform feature fusion and vector-level batch normalization on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain the final fused feature vector. The processing logic includes: Perform vector outer product operation on the power modal feature embedding vector and the thermal modal feature embedding vector to obtain the modal interaction tensor, and perform Z-score normalization and vector splicing to obtain the combined input vector, which is expressed as follows: Combined input vector = [electrical modal feature embedding vector, thermal modal feature embedding vector, modal interaction tensor]; The InTense fusion network is used to perform feature fusion and batch normalization on the combined input vector. The processing logic includes: The combined input vector is multiplied by the first weight matrix in the linear combination layer, added to the first bias term, and then nonlinearly transformed by the ReLU function to obtain the global fusion feature. The dimension of the first weight matrix is ​​256×1088. The modal interaction tensor is multiplied by the second weight matrix in the tensor product interaction layer, added to the second bias term, and then nonlinearly transformed by the ReLU function to obtain the interaction reinforcement feature. The dimension of the second weight matrix is ​​256×1024. The global fusion features and interactive enhancement features are added element by element to obtain the preliminary fusion features, and the preliminary fusion features are batch normalized by the vector-level batch normalization (VBN) algorithm to obtain the final fusion feature vector.

7. The coordinated control method of a multi-modal energy storage system according to claim 1, characterized in that: Step S3: Calculate the contribution based on the final fused feature vector to obtain a modal contribution ratio set. The processing logic includes: The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the power modal feature embedding vector to obtain a first sub-matrix, where the dimension of the first sub-matrix is ​​256×32, and the 32 dimensions corresponding to the power modal feature embedding vector are the first 32 columns of the combined input vector; The square root of the sum of squares of the elements in each row of the first submatrix is ​​calculated to obtain the first L2 norm set, and the arithmetic mean of the corresponding norms of each row in the first L2 norm set is taken to obtain the original value of the power modal contribution; The first weight matrix is ​​truncated based on the 32 dimensions corresponding to the thermal modal feature embedding vector to obtain a second sub-matrix, where the dimension of the second sub-matrix is ​​256×32, and the 32 dimensions corresponding to the thermal modal feature embedding vector are the 33rd to 64th columns of the combined input vector; The square root of the sum of squares of the elements in each row of the second submatrix is ​​calculated to obtain the second L2 norm set, and the arithmetic mean of the corresponding norms of each row in the second L2 norm set is taken to obtain the original value of the thermal modal contribution; The square root of the sum of squares of the elements in each row of the second weight matrix is ​​calculated to obtain the third L2 norm set, and the arithmetic mean of the corresponding norms of each row in the third L2 norm set is taken to obtain the original value of the modal interaction contribution; A modal contribution ratio set is obtained by summing and normalizing the original values ​​of the power modal contribution, the thermal modal contribution, and the modal interaction contribution. The modal contribution ratio set includes the power modal contribution ratio, the thermal modal contribution ratio, and the modal interaction contribution ratio. The calculation logic of the summation and normalization includes: Normalized base value = original value of power modal contribution + original value of thermal modal contribution + original value of modal interaction contribution; Power modal contribution ratio = power modal contribution original value / normalized base value; Thermal mode contribution ratio = thermal mode contribution original value / normalized base value; Modal interaction contribution ratio = original value of modal interaction contribution / normalized base value.

8. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: Step S4: Establish a linear regression model and calculate the predicted load data using the weighted least squares method. The processing logic includes: The final fusion feature vector, external features and historical loads are Z-score normalized and vector concatenated to obtain linear regression input items and establish a linear regression model; The external features include real-time electricity prices and real-time weather temperatures; The historical load includes historical power load data and historical thermal load data within 14 days, divided into 5-minute steps. The historical power load data includes historical total active load values, and the historical thermal load data includes historical cooling load values. The calculation expression of the linear regression model is: ; Where t represents time, P(t) represents the predicted load data, represents the linear regression input, represents the final fusion feature vector, D(t) represents the external feature, H(t) represents the historical load, represents the linear regression weight matrix, represents the linear regression bias term; The linear regression model is fitted and calculated by weighted least squares (WLS) to obtain the predicted load data for the next 6 hours. The predicted load data includes predicted power load data and predicted thermal load data, which are divided into 72 time steps with a step length of 5 minutes.

9. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: Step S5: Combine the modal contribution ratio set and the predicted load data to perform particle swarm optimization calculation to obtain the energy storage charging and discharging strategy. The processing logic includes: With the goal of minimizing the electricity purchase cost, a penalty term is constructed based on the power mode contribution ratio and the thermal mode contribution ratio, and a mixed integer quadratic programming (MIQP) model is established. Its calculation expression is: ； in, represents the minimum electricity purchase cost, U(t) represents the electricity purchase price, represents the purchased power, λ represents the penalty coefficient, represents the power mode contribution ratio, represents the power forecast error, represents the contribution ratio of thermal mode, represents the thermal prediction error; The mixed integer quadratic programming model is optimized by combining the APSO particle swarm optimization algorithm with the modal interaction contribution ratio. The processing logic includes: The inertia weight is calculated by linear mapping based on the modal interaction contribution ratio. The speed update formula of the particle swarm optimization algorithm is adjusted by the inertia weight to obtain the particle swarm search parameter. The energy storage charging and discharging strategy is obtained by iterative calculation based on the particle swarm search parameter. The energy storage charging and discharging strategy includes the battery charging and discharging power and SOC adjustment target, divided into 5-minute steps, with a total of 72 time steps; When the battery charge and discharge power is a positive number, it indicates charging, and when it is a negative number, it indicates discharging; The calculation expression of the inertia weight is: ； in, represents the inertia weight, represents the initial global search weight, represents the local convergence weight, Indicates the contribution ratio of modal interaction.

10. The coordinated control method of a multi-modal energy storage system according to claim 1, wherein: In step S6, the energy storage charge and discharge strategy is converted into JSON format to obtain control instruction data, and the control instruction data is sent to the energy storage system via the HTTP protocol. The control instruction data is then transmitted to the battery management system (BMS) and the power storage converter (PCS) on the energy storage system device side via the CAN protocol. The battery management system and the power storage converter adjust the charge and discharge current based on the control instruction data to complete real-time control.

Images