Intelligent comprehensive power box power supply optimization method and system
By using the multi-dimensional sensing array and data processing technology of the intelligent integrated power box, combined with LSTM and SVR models and Nash equilibrium algorithm, the problems of data quality and prediction accuracy of traditional power boxes are solved, and efficient and reliable power supply optimization is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG YONGFENGSHUO ELECTRIC POWER INTEGRATION TECH CO LTD
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional power box control strategies face problems such as poor data quality, low prediction accuracy, difficulty in optimizing scheduling, and weak adaptive capabilities. In particular, when dealing with high-dimensional and highly correlated power system data, it is difficult to accurately characterize load fluctuation characteristics and the uncertainty of new energy sources, and there is a lack of multi-entity collaborative optimization mechanisms.
Data is collected in real time using a multi-dimensional sensing array, and denoising is performed by combining wavelet transform and Kalman filtering. After normalization using a sliding window, key latent features are extracted by principal component analysis. An LSTM and SVR fusion model is constructed for prediction, and a Nash equilibrium algorithm is introduced to optimize the power allocation strategy of the source-grid-load-storage unit.
It significantly improves the efficiency and reliability of power supply. Through high-quality data processing and dynamic optimization, it accurately captures load and new energy fluctuations, achieves economical operation with the lowest life cycle cost, and ensures voltage stability and load balance.
Smart Images
Figure CN122159294A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power supply technology for power boxes, specifically to a method and system for optimizing power supply in intelligent integrated power boxes. Background Technology
[0002] With the large-scale integration of distributed energy resources (such as photovoltaic and wind power) and the increase in fluctuating loads such as electric vehicles and sensitive electronic devices, the power supply and consumption environment at the end of the distribution network is becoming increasingly complex. As a key node connecting the distribution network and the user side, the intelligent integrated power box undertakes important functions of power distribution, metering, protection, and optimized supply. However, in actual operation, traditional power box control strategies face multiple technical challenges, including poor data quality, low prediction accuracy, difficulty in optimized scheduling, and weak adaptive capabilities.
[0003] First, in terms of feature extraction and state awareness, traditional methods typically use explicit physical quantities such as voltage and current for control, ignoring the implicit features behind the data. Because power system data is highly dimensional and strongly correlated, directly using the entire dataset is not only computationally burdensome but also prone to the "curse of dimensionality." Existing feature processing methods lack the ability to extract key implicit features such as "load volatility," "new energy output confidence," and "load importance" using algorithms like principal component analysis, making it difficult for the system to accurately characterize the random fluctuations of the load and the uncertainty of new energy sources.
[0004] Furthermore, in terms of power forecasting and trend analysis, a single forecasting model struggles to simultaneously account for the long-term dependence and nonlinear fluctuation characteristics of time series data. Existing power distribution boxes mostly employ basic neural networks or statistical regression methods, which cannot effectively output power forecast curves with confidence intervals. This results in a lack of robustness in the forecasting results when faced with sudden load changes or weather fluctuations, failing to provide reliable probabilistic boundaries for subsequent optimized scheduling.
[0005] Furthermore, in terms of optimizing scheduling and control strategies, existing control methods are mostly based on deterministic rules or single objectives (such as considering only voltage stability), lacking a comprehensive consideration of lifecycle costs, voltage deviation, and load balance. More importantly, the relationships between the source, grid, load, and storage units are essentially one of interest games. Existing technologies lack mechanisms to introduce game theory (such as Nash equilibrium) to solve multi-agent collaborative optimization problems, making it difficult to achieve a balance of interests among units and global optimization within a non-cooperative game framework. Summary of the Invention
[0006] To address the aforementioned technical problems, this paper provides a method and system for optimizing power supply in an intelligent integrated power box. This technical solution resolves the issues raised in the background section.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, a method for optimizing power supply in an intelligent integrated power box is provided, comprising: By deploying a multi-dimensional sensing array in the intelligent integrated power box, real-time data on power supply side, load side, energy storage status, and environmental parameters are collected to form a raw time-series dataset. Edge computing preprocessing is performed on the original time series dataset. A denoising algorithm combining wavelet transform and Kalman filtering is used to remove outliers and noise. A sliding window mechanism is used to perform time alignment and normalization on the data. Based on the preprocessed data, a multidimensional feature vector is constructed, and the principal component analysis algorithm is used to reduce the dimensionality of the features and extract the key implicit features that affect the efficiency of power supply, including load volatility, confidence of new energy output, and load importance factor. A prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR) is established to perform rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and output power prediction curves with confidence intervals. A multi-objective optimization function is constructed with the objectives of minimizing the total life cycle cost, minimizing voltage deviation, and maximizing load balance. The Nash equilibrium algorithm from game theory is introduced to solve the power allocation strategy among the source-grid-load-storage units. Based on the power allocation strategy obtained from the solution, specific action commands are generated for the intelligent circuit breaker, bidirectional converter and flexible load controller, and then sent to the actuator via industrial Ethernet.
[0008] Preferably, the denoising algorithm combining wavelet transform and Kalman filtering specifically includes: First, the original time series data is decomposed into multiple scales using wavelet transform. Then, the db4 wavelet basis function is selected to perform soft threshold quantization on the high-frequency noise coefficients while retaining the low-frequency approximation coefficients. The reconstructed low-frequency signal was then used as the observation value for Kalman filtering to establish a linear discrete system model based on the state transition matrix. By iteratively updating the prediction error covariance matrix, peak outliers caused by instantaneous sensor failures were eliminated. The formula for calculating the Kalman gain is as follows: ; in, The Kalman gain at time k; For the first The prior estimate error covariance matrix at time t; The observation matrix; The observation noise variance matrix; The state update equation for the Kalman filter is: ; in, For the first The posterior state estimate at time t; For the first The prior state estimate at time t; For the first The observed value at that moment.
[0009] Preferably, the step of using a sliding window mechanism to perform time alignment and normalization processing on the data specifically includes: Set a fixed-length sliding window, with the window length set to 50ms to 200ms based on the response delay characteristics of the power box, and the window overlap rate set to 50%; Within the window, the data is normalized using the Z-Score normalization method, and linear interpolation is used to fill in the microsecond-level timestamp loss caused by communication packet loss. The standardized Z-Score calculation formula is as follows: ; in, The data is after normalization; These are the original data points; This is the mean of all data within the current sliding window; This represents the standard deviation of all data within the current sliding window. The formula for linear interpolation is: ; in, The data values to be filled in by interpolation; Timestamps for missing data; , The timestamp and value of the previous valid data point before the missing point; , This is the timestamp and value of the next valid data point after the missing point.
[0010] Preferably, the feature dimensionality reduction using principal component analysis specifically includes: Construct a standardized feature matrix X containing voltage, current, frequency, temperature, and irradiance, and calculate the eigenvalues and eigenvectors of the covariance matrix; sort the eigenvalues from largest to smallest, and select the top k principal components with a cumulative variance contribution rate greater than 95%; The formula for calculating the covariance matrix is: ; Where Σ is the covariance matrix; The number of samples; The data matrix after centralization; The formula for calculating the cumulative variance contribution rate is: ; in, For the front Cumulative variance contribution rate of each principal component; For the first One eigenvalue; This represents the total dimension of the original features.
[0011] Preferably, the extraction of key latent features affecting power supply efficiency specifically includes: Load volatility is expressed using the coefficient of variation, and the calculation formula is as follows: ; in, For load volatility; The standard deviation of the load power within the sliding window; This represents the average load power within the sliding window; The confidence level of new energy output is calculated based on the probability distribution fitting of historical output data, using the Beta distribution function. Specifically: ; in, Confidence in contributing to new energy sources; , These represent the upper and lower limits of the power output prediction range under current meteorological conditions. For parameters The probability density function of the Beta distribution; The load importance factor is calculated by constructing a judgment matrix using the analytic hierarchy process (AHP) and combining the weights of load outage loss cost and recovery time. ; in, This is the load importance factor; , These are the weights for power outage loss costs and recovery time, respectively. Economic loss per unit of power outage time; The time required to restore power to the load.
[0012] Preferably, the establishment of a prediction model based on the fusion of Long Short-Term Memory (LSTM) network and Support Vector Regression (SVR) machine specifically includes: First, the preprocessed time series data is input into the LSTM network, which uses its gating mechanism to capture long-term dependencies in the time series and outputs a preliminary predicted trend vector. This trend vector is then used as the input feature of the support vector regression machine (SVR), and the radial basis function (RBF) is used to map the data to a high-dimensional feature space. The regression function expression for SVR is: ; in, To predict the output value; The number of support vectors; , For Lagrange multipliers; For bias terms; The radial basis function kernel is calculated using the following formula: ; in, This is the width parameter of the kernel function; is the Euclidean distance between the input vector and the support vectors.
[0013] Preferably, the output power prediction curve with confidence intervals specifically includes: Using quantile regression, regression models were constructed for the upper quantile of 95% and the lower quantile of 5%, respectively. A quantile loss function was defined, and the model was trained by minimizing the sum of weighted absolute errors between the predicted and the true values. The formula for calculating the quantile loss function is: ; in, Quantile loss; This is the actual power value; To predict power values; For quantile levels, take 0.05 or 0.95; The formula for calculating the final output confidence interval width is: ; in, The width of the confidence interval; This is the critical value for the standard normal distribution; This represents the standard deviation of the historical prediction residuals.
[0014] Preferably, the mathematical expression for constructing the multi-objective optimization function is: ; in, The value of the comprehensive objective function; , , The dynamic weighting coefficients satisfy the following conditions: + + ; Total life cycle cost The calculation formula is: ; in, This refers to the initial investment cost; For the first Annual maintenance costs; For the first Annual replacement cost; The discount rate; Lifespan (in years); Voltage deviation The calculation formula is: ; in, This represents the actual voltage at the node. Rated voltage; Load balance index The calculation formula is: ; in, Three-phase load current The variance.
[0015] Preferably, the specific iterative process of introducing the Nash equalization algorithm to solve the power allocation strategy is as follows: By treating each unit—source, network, load, and storage—as a game participant, constructing their respective utility functions, and using the gradient descent method within a non-cooperative game framework, we can solve for the optimal response function of each unit given the strategies of other units. The utility function of a power generation unit is defined as: ; in, For the utility of the power generation unit; For revenue from electricity sales; The cost of electricity generation includes fuel and maintenance costs. The utility function of an energy storage unit is defined as: ; in, For the utility of energy storage units; For peak-valley arbitrage profits; Costs associated with charging and discharging losses; The criterion for determining the Nash equilibrium is: for any participant ,satisfy: ; in, For participants Equilibrium strategy; For the equilibrium strategy combination of other participants; For participants Any feasible strategy; when the strategy deviation of all units is less than the set threshold for three consecutive iterations, it is determined that Nash equilibrium has been reached.
[0016] In a second aspect of the invention, an intelligent integrated power box power supply optimization system is also provided, comprising: The acquisition module is used to collect power supply side data, load side data, energy storage status data and environmental parameter data in real time through a multi-dimensional sensing array deployed in the intelligent integrated power box, forming a raw time series dataset. The preprocessing module is used to perform edge computing preprocessing on the original time series dataset, and uses a denoising algorithm combining wavelet transform and Kalman filtering to remove outliers and noise, and uses a sliding window mechanism to perform time alignment and normalization on the data. The extraction module is used to construct a multi-dimensional feature vector based on the preprocessed data, perform feature dimensionality reduction using principal component analysis algorithm, and extract key implicit features affecting power supply efficiency, including load volatility, new energy output confidence and load importance factor. The first establishment module is used to establish a prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR), and to perform rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and output a power prediction curve with confidence intervals. The second establishment module is used to construct a multi-objective optimization function with the optimization objectives of the lowest life cycle cost, the smallest voltage deviation, and the highest load balance. The Nash equilibrium algorithm in game theory is introduced to solve the power allocation strategy among the source-grid-load-storage units. The allocation module is used to generate specific action instructions for the intelligent circuit breaker, bidirectional converter and flexible load controller according to the power allocation strategy obtained by solving, and send them to the actuator via industrial Ethernet.
[0017] Compared with the prior art, the present invention provides a method and system for optimizing power supply in an intelligent integrated power box, which has the following beneficial effects: This invention significantly improves the power supply efficiency and reliability of intelligent integrated power boxes through refined end-to-end data processing and dynamic optimization. First, it employs a denoising algorithm combining wavelet transform and Kalman filtering, along with a sliding window normalization mechanism, to effectively eliminate sensor noise and outliers, solving the time alignment problem of multi-source heterogeneous data and ensuring high quality and consistency of the original data. Second, it utilizes the PCA algorithm to extract key implicit features such as load volatility and renewable energy output confidence, achieving dimensionality reduction and deep state perception of high-dimensional data. Combining an LSTM and SVR fusion model, it outputs power prediction curves with confidence intervals, accurately capturing the nonlinear fluctuations of load and renewable energy, significantly improving prediction robustness. Third, it constructs a multi-objective optimization function encompassing lifecycle cost, voltage deviation, and load balance, introducing a Nash equilibrium algorithm to solve the power allocation strategy for each unit (source-grid-load-storage), achieving the economical operation goal of lowest lifecycle cost while ensuring voltage stability and load balance. Finally, it precisely issues control commands via industrial Ethernet to achieve coordinated action of various actuators, effectively solving the problems of poor data quality, low prediction accuracy, and difficult optimization scheduling in traditional power boxes. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the power supply optimization method for intelligent integrated power boxes in this invention; Figure 2 This is a schematic diagram of the framework of the intelligent integrated power box power supply optimization system in this invention. Detailed Implementation
[0019] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0020] Example 1 Please refer to Figure 1 As shown, in a first aspect of the present invention, a method for optimizing power supply in an intelligent integrated power box is provided, comprising: S101. Through the multi-dimensional sensing array deployed in the intelligent integrated power box, real-time data of power supply side, load side, energy storage status and environmental parameters are collected to form the original time series dataset. S102. Perform edge computing preprocessing on the original time series dataset, use a denoising algorithm combining wavelet transform and Kalman filtering to remove outliers and noise, and use a sliding window mechanism to perform time alignment and normalization on the data. S103. Based on the preprocessed data, construct a multidimensional feature vector, use principal component analysis algorithm to perform feature dimensionality reduction, and extract key implicit features that affect power supply efficiency, including load volatility, new energy output confidence and load importance factor. S104. Establish a prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR) to make rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and output a power prediction curve with confidence intervals. S105. Construct a multi-objective optimization function with the objectives of minimizing the total life cycle cost, minimizing the voltage deviation, and maximizing the load balance. Introduce the Nash equilibrium algorithm from game theory to solve the power allocation strategy among the source-grid-load-storage units. S106. Based on the power allocation strategy obtained from the solution, generate specific action commands for the intelligent circuit breaker, bidirectional converter and flexible load controller, and send them to the actuator via industrial Ethernet.
[0021] As will be understood by those skilled in the art, this invention significantly improves the power supply efficiency and reliability of intelligent integrated power boxes through refined processing and dynamic optimization of end-to-end data. First, it employs a denoising algorithm combining wavelet transform and Kalman filtering, along with a sliding window normalization mechanism, to effectively eliminate sensor noise and outliers, solving the time alignment problem of multi-source heterogeneous data and ensuring high quality and consistency of the original data. Second, it utilizes the PCA algorithm to extract key implicit features such as load volatility and new energy output confidence, achieving dimensionality reduction and deep state perception of high-dimensional data. Combining the LSTM and SVR fusion model, it outputs power prediction curves with confidence intervals, accurately capturing the nonlinear fluctuations of load and new energy, and significantly improving prediction robustness. Third, it constructs a multi-objective optimization function encompassing lifecycle cost, voltage deviation, and load balance, introducing a Nash equilibrium algorithm to solve the power allocation strategy for each unit (source-grid-load-storage), achieving the economical operation objective of lowest lifecycle cost while ensuring voltage stability and load balance. Finally, by precisely issuing control commands via industrial Ethernet, coordinated actions of various actuators are achieved, effectively solving the problems of poor data quality, low prediction accuracy, and difficulty in optimizing scheduling in traditional power boxes.
[0022] The denoising algorithm that combines wavelet transform and Kalman filtering is as follows: First, the original time series data is decomposed into multiple scales using wavelet transform. Then, the db4 wavelet basis function is selected to perform soft threshold quantization on the high-frequency noise coefficients while retaining the low-frequency approximation coefficients. The reconstructed low-frequency signal was then used as the observation value for Kalman filtering to establish a linear discrete system model based on the state transition matrix. By iteratively updating the prediction error covariance matrix, peak outliers caused by instantaneous sensor failures were eliminated. The formula for calculating the Kalman gain is as follows: ; in, The Kalman gain at time k; For the first The prior estimate error covariance matrix at time t; The observation matrix; The observation noise variance matrix; The state update equation for the Kalman filter is: ; in, For the first The posterior state estimate at time t; For the first The prior state estimate at time t; For the first The observed value at that moment.
[0023] The sliding window mechanism is used to perform time alignment and normalization of the data, specifically including: Set a fixed-length sliding window, with the window length set to 50ms to 200ms based on the response delay characteristics of the power box, and the window overlap rate set to 50%; Within the window, the data is normalized using the Z-Score normalization method, and linear interpolation is used to fill in the microsecond-level timestamp loss caused by communication packet loss. The standardized formula for AA-Score is as follows: ; in, The data is after normalization; These are the original data points; This is the mean of all data within the current sliding window; This represents the standard deviation of all data within the current sliding window. The formula for linear interpolation is: ; in, The data values to be filled in by interpolation; Timestamps for missing data; , The timestamp and value of the previous valid data point before the missing point; , This is the timestamp and value of the next valid data point after the missing point.
[0024] Feature dimensionality reduction using principal component analysis (PCA) algorithms specifically includes: Construct a standardized feature matrix X containing voltage, current, frequency, temperature, and irradiance, and calculate the eigenvalues and eigenvectors of the covariance matrix; Sort the eigenvalues from largest to smallest and select the top k principal components with a cumulative variance contribution rate greater than 95%. The formula for calculating the covariance matrix is: ; Where Σ is the covariance matrix; The number of samples; The data matrix after centralization; The formula for calculating the cumulative variance contribution rate is: ; in, For the front Cumulative variance contribution rate of each principal component; For the first One eigenvalue; This represents the total dimension of the original features.
[0025] Extracting key implicit features affecting power supply efficiency, specifically including: Load volatility is expressed using the coefficient of variation, and the calculation formula is as follows: ; in, For load volatility; The standard deviation of the load power within the sliding window; This represents the average load power within the sliding window; The confidence level of new energy output is calculated based on the probability distribution fitting of historical output data, using the Beta distribution function. Specifically: ; in, Confidence in contributing to new energy sources; , These represent the upper and lower limits of the power output prediction range under current meteorological conditions. For parameters The probability density function of the Beta distribution; The load importance factor is calculated by constructing a judgment matrix using the analytic hierarchy process (AHP) and combining the weights of load outage loss cost and recovery time. ; in, This is the load importance factor; , These are the weights for power outage loss costs and recovery time, respectively. Economic loss per unit of power outage time; The time required to restore power to the load.
[0026] A prediction model based on the fusion of Long Short-Term Memory (LSTM) network and Support Vector Regression (SVR) machine is established, specifically including: First, the preprocessed time series data is input into the LSTM network, which uses its gating mechanism to capture long-term dependencies in the time series and outputs a preliminary predicted trend vector. This trend vector is then used as the input feature of the support vector regression machine (SVR), and the radial basis function (RBF) is used to map the data to a high-dimensional feature space. The regression function expression for SVR is: ; in, To predict the output value; The number of support vectors; , For Lagrange multipliers; For bias terms; The radial basis function kernel is calculated using the following formula: ; in, This is the width parameter of the kernel function; is the Euclidean distance between the input vector and the support vectors.
[0027] Output power prediction curves with confidence intervals, specifically including: Using quantile regression, regression models were constructed for the upper quantile of 95% and the lower quantile of 5%, respectively. A quantile loss function was defined, and the model was trained by minimizing the sum of weighted absolute errors between the predicted and the true values. The formula for calculating the quantile loss function is: ; in, Quantile loss; This is the actual power value; To predict power values; For quantile levels, take 0.05 or 0.95; The formula for calculating the final output confidence interval width is: ; in, The width of the confidence interval; This is the critical value for the standard normal distribution; This represents the standard deviation of the historical prediction residuals.
[0028] The mathematical expression for constructing the multi-objective optimization function is as follows: ; in, The value of the comprehensive objective function; , , The dynamic weighting coefficients satisfy the following conditions: + + ; Total life cycle cost The calculation formula is: ; in, This refers to the initial investment cost; For the first Annual maintenance costs; For the first Annual replacement cost; The discount rate; Lifespan (in years); Voltage deviation The calculation formula is: ; in, This represents the actual voltage at the node. Rated voltage; Load balance index The calculation formula is: ; in, Three-phase load current The variance.
[0029] The specific iterative process of using the Nash equalization algorithm to solve the power allocation strategy is as follows: By treating each unit—source, network, load, and storage—as a game participant, constructing their respective utility functions, and using the gradient descent method within a non-cooperative game framework, we can solve for the optimal response function of each unit given the strategies of other units. The utility function of a power generation unit is defined as: ; in, For the utility of the power generation unit; For revenue from electricity sales; The cost of electricity generation includes fuel and maintenance costs. The utility function of an energy storage unit is defined as: ; in, For the utility of energy storage units; For peak-valley arbitrage profits; Costs associated with charging and discharging losses; The criterion for determining the Nash equilibrium is: for any participant ,satisfy: ; in, For participants Equilibrium strategy; For the equilibrium strategy combination of other participants; For participants Any feasible strategy; when the strategy deviation of all units is less than the set threshold for three consecutive iterations, it is determined that Nash equilibrium has been reached.
[0030] Please refer to Figure 2 As shown, in a second aspect of the present invention, an intelligent integrated power box power supply optimization system is also provided, the system comprising: The data acquisition module is used to collect real-time data on the power supply side, load side, energy storage status, and environmental parameters through a multi-dimensional sensing array deployed in the intelligent integrated power box, forming a raw time-series dataset. The preprocessing module is used to perform edge computing preprocessing on the original time series dataset. It uses a denoising algorithm combining wavelet transform and Kalman filtering to remove outliers and noise, and uses a sliding window mechanism to perform time alignment and normalization on the data. The extraction module is used to construct a multi-dimensional feature vector based on the preprocessed data, and to perform feature dimensionality reduction using the principal component analysis algorithm to extract key implicit features that affect the efficiency of power supply, including load volatility, confidence of new energy output, and load importance factor. The first module is used to build a prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR). It performs rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and outputs a power prediction curve with confidence intervals. The second module is used to construct a multi-objective optimization function with the objectives of minimizing the total life cycle cost, minimizing the voltage deviation, and maximizing the load balance. It introduces the Nash equilibrium algorithm from game theory to solve the power allocation strategy among the source-grid-load-storage units. The power allocation module generates specific action commands for the intelligent circuit breaker, bidirectional converter, and flexible load controller based on the power allocation strategy obtained from the solution, and sends them to the actuators via industrial Ethernet.
[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for optimizing power supply in an intelligent integrated power box, characterized in that, include: By deploying a multi-dimensional sensing array in the intelligent integrated power box, real-time data on power supply side, load side, energy storage status, and environmental parameters are collected to form a raw time-series dataset. Edge computing preprocessing is performed on the original time series dataset. A denoising algorithm combining wavelet transform and Kalman filtering is used to remove outliers and noise. A sliding window mechanism is used to perform time alignment and normalization on the data. Based on the preprocessed data, a multidimensional feature vector is constructed, and the principal component analysis algorithm is used to reduce the dimensionality of the features and extract the key implicit features that affect the efficiency of power supply, including load volatility, confidence of new energy output, and load importance factor. A prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR) is established to perform rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and output power prediction curves with confidence intervals. A multi-objective optimization function is constructed with the objectives of minimizing the total life cycle cost, minimizing voltage deviation, and maximizing load balance. The Nash equilibrium algorithm from game theory is introduced to solve the power allocation strategy among the source-grid-load-storage units. Based on the power allocation strategy obtained from the solution, specific action commands are generated for the intelligent circuit breaker, bidirectional converter and flexible load controller, and then sent to the actuator via industrial Ethernet.
2. The method for optimizing power supply in an intelligent integrated power box according to claim 1, characterized in that, The denoising algorithm that combines wavelet transform and Kalman filtering is as follows: First, the original time series data is decomposed into multiple scales using wavelet transform. Then, the db4 wavelet basis function is selected to perform soft threshold quantization on the high-frequency noise coefficients while retaining the low-frequency approximation coefficients. The reconstructed low-frequency signal was then used as the observation value for Kalman filtering to establish a linear discrete system model based on the state transition matrix. By iteratively updating the prediction error covariance matrix, peak outliers caused by instantaneous sensor failures were eliminated. The formula for calculating the Kalman gain is as follows: ； in, The Kalman gain at time k; For the first The prior estimate error covariance matrix at time t; The observation matrix; The observation noise variance matrix; The state update equation for the Kalman filter is: ； in, For the first The posterior state estimate at time t; For the first The prior state estimate at time t; For the first The observed value at that moment.
3. The method for optimizing power supply in an intelligent integrated power box according to claim 2, characterized in that, The process of using a sliding window mechanism to perform time alignment and normalization of data specifically includes: Set a fixed-length sliding window, with the window length set to 50ms to 200ms based on the response delay characteristics of the power box, and the window overlap rate set to 50%; Within the window, the data is normalized using the Z-Score normalization method, and linear interpolation is used to fill in the microsecond-level timestamp loss caused by communication packet loss. The standardized Z-Score calculation formula is as follows: ； in, The data is after normalization; These are the original data points; This is the mean of all data within the current sliding window; This represents the standard deviation of all data within the current sliding window. The formula for linear interpolation is: ； in, The data values to be filled in by interpolation; Timestamps for missing data; , The timestamp and value of the previous valid data point before the missing point; , This is the timestamp and value of the next valid data point after the missing point.
4. The method for optimizing power supply in an intelligent integrated power box according to claim 3, characterized in that, The feature dimensionality reduction using principal component analysis specifically includes: Construct a standardized feature matrix X containing voltage, current, frequency, temperature, and irradiance, and calculate the eigenvalues and eigenvectors of the covariance matrix; sort the eigenvalues from largest to smallest, and select the top k principal components with a cumulative variance contribution rate greater than 95%; The formula for calculating the covariance matrix is: ； Where Σ is the covariance matrix; The number of samples; The data matrix after centralization; The formula for calculating the cumulative variance contribution rate is: ； in, For the front Cumulative variance contribution rate of each principal component; For the first One eigenvalue; This represents the total dimension of the original features.
5. The method for optimizing power supply in an intelligent integrated power box according to claim 4, characterized in that, The extraction of key latent features affecting power supply efficiency specifically includes: Load volatility is expressed using the coefficient of variation, and the calculation formula is as follows: ； in, For load volatility; The standard deviation of the load power within the sliding window; This represents the average load power within the sliding window; The confidence level of new energy output is calculated based on the probability distribution fitting of historical output data, using the Beta distribution function. Specifically: ； in, Confidence in contributing to new energy sources; , These represent the upper and lower limits of the power output prediction range under current meteorological conditions. For parameters The probability density function of the Beta distribution; The load importance factor is calculated by constructing a judgment matrix using the analytic hierarchy process (AHP) and combining the weights of load outage loss cost and recovery time. ； in, This is the load importance factor; , These are the weights for power outage loss costs and recovery time, respectively. Economic loss per unit of power outage time; The time required to restore power to the load.
6. The method for optimizing power supply in an intelligent integrated power box according to claim 5, characterized in that, The establishment of the prediction model based on the fusion of Long Short-Term Memory (LSTM) network and Support Vector Regression (SVR) machine specifically includes: First, the preprocessed time series data is input into the LSTM network, which uses its gating mechanism to capture long-term dependencies in the time series and outputs a preliminary predicted trend vector. This trend vector is then used as the input feature of the support vector regression machine (SVR), and the radial basis function (RBF) is used to map the data to a high-dimensional feature space. The regression function expression for SVR is: ； in, To predict the output value; The number of support vectors; , For Lagrange multipliers; For bias terms; The radial basis function kernel is calculated using the following formula: ； in, This is the width parameter of the kernel function; is the Euclidean distance between the input vector and the support vectors.
7. The method for optimizing power supply in an intelligent integrated power box according to claim 6, characterized in that, The output power prediction curve with confidence intervals specifically includes: Using quantile regression, regression models were constructed for the upper quantile of 95% and the lower quantile of 5%, respectively. A quantile loss function was defined, and the model was trained by minimizing the sum of weighted absolute errors between the predicted and the true values. The formula for calculating the quantile loss function is: ； in, Quantile loss; This is the actual power value; To predict power values; For quantile levels, take 0.05 or 0.95; The formula for calculating the final output confidence interval width is: ； in, The width of the confidence interval; This is the critical value for the standard normal distribution; This represents the standard deviation of the historical prediction residuals.
8. The method for optimizing power supply in an intelligent integrated power box according to claim 7, characterized in that, The mathematical expression for constructing the multi-objective optimization function is as follows: ； in, The value of the comprehensive objective function; , , These are dynamic weighting coefficients, and they satisfy... + + ; Total life cycle cost The calculation formula is: ； in, This refers to the initial investment cost; For the first Annual maintenance costs; For the first Annual replacement cost; The discount rate; Lifespan (in years); Voltage deviation The calculation formula is: ； in, This represents the actual voltage at the node. Rated voltage; Load balance index The calculation formula is: ； in, Three-phase load current The variance.
9. The method for optimizing power supply in an intelligent integrated power box according to claim 8, characterized in that, The specific iterative process of introducing the Nash equalization algorithm to solve the power allocation strategy is as follows: By treating each unit—source, network, load, and storage—as a game participant, constructing their respective utility functions, and using the gradient descent method within a non-cooperative game framework, we can solve for the optimal response function of each unit given the strategies of other units. The utility function of a power generation unit is defined as: ； in, For the utility of the power generation unit; For revenue from electricity sales; The cost of electricity generation includes fuel and maintenance costs. The utility function of an energy storage unit is defined as: ； in, For the utility of energy storage units; For peak-valley arbitrage profits; Costs associated with charging and discharging losses; The criterion for determining the Nash equilibrium is: for any participant ,satisfy: ； in, For participants Equilibrium strategy; For the equilibrium strategy combination of other participants; For participants Any feasible strategy; when the strategy deviation of all units is less than the set threshold for three consecutive iterations, it is determined that Nash equilibrium has been reached.
10. A smart integrated power box power supply optimization system, used to implement the smart integrated power box power supply optimization method as described in any one of claims 1-9, characterized in that, include: The acquisition module is used to collect power supply side data, load side data, energy storage status data and environmental parameter data in real time through a multi-dimensional sensing array deployed in the intelligent integrated power box, forming a raw time series dataset. The preprocessing module is used to perform edge computing preprocessing on the original time series dataset, and uses a denoising algorithm combining wavelet transform and Kalman filtering to remove outliers and noise, and uses a sliding window mechanism to perform time alignment and normalization on the data. The extraction module is used to construct a multi-dimensional feature vector based on the preprocessed data, perform feature dimensionality reduction using principal component analysis algorithm, and extract key implicit features that affect the efficiency of power supply, including load volatility, confidence level of new energy output, and load importance factor. The first establishment module is used to establish a prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and Support Vector Regression Machine (SVR), and to perform rolling predictions of load demand power and photovoltaic / wind power output power within a set future time window, and output a power prediction curve with confidence intervals. The second establishment module is used to construct a multi-objective optimization function with the optimization objectives of the lowest life cycle cost, the smallest voltage deviation, and the highest load balance. The Nash equilibrium algorithm in game theory is introduced to solve the power allocation strategy among the source-grid-load-storage units. The allocation module is used to generate specific action instructions for the intelligent circuit breaker, bidirectional converter and flexible load controller according to the power allocation strategy obtained by solving, and send them to the actuator via industrial Ethernet.