UPS energy storage system and method based on power management

Abstract

The present invention relates to the technical field of energy storage management, and specifically discloses a UPS energy storage system and method based on power management. The system collects electrical status parameters of each UPS node in a cluster in real time through a parallel monitoring module, collects status parameters of the power grid end through a power grid analyzer, performs load characteristic fluctuation phase analysis on the denoised UPS node characteristic data set, and extracts the time domain load change gradient of each UPS node; predicts the load demand range of each UPS node in the next 15 minutes through an LSTM timing network; generates a dynamic energy efficiency MAP diagram of each UPS node based on historical energy efficiency curve data, and calibrates the optimal energy efficiency working range of each UPS node in combination with real-time SOC; generates a cluster-level load distribution feasibility matrix based on a cross-comparison between the load demand range and the optimal energy efficiency working range; and generates a cluster load distribution control instruction set based on the matrix to ensure a smooth transition of each UPS node during load switching, reduce the impact caused by the switching, and improve the overall efficiency and safety of the system.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy storage management, and relates to a UPS energy storage system and method based on power management. Background Art

[0002] A UPS (Uninterruptible Power Supply) is an uninterruptible power supply (UPS) that includes an energy storage device. In today's digital age, critical facilities such as data centers and communication base stations are highly dependent on UPSs. The waste of conversion efficiency caused by the low load factor of a single UPS has become a core issue that needs to be addressed urgently.

[0003] In traditional UPS cluster management, the problem of efficiency loss caused by low load factors has long existed (for example, the efficiency of a single UPS drops sharply when the load falls below 40%). Existing technologies rely on static load balancing or simple module sleep strategies, making it difficult to adapt to dynamic load fluctuations and complex grid disturbance scenarios. For example, while traditional parallel redundant UPS systems (such as centralized bypass architectures) have redundancy, they lack the ability to accurately perceive and control node-level electrical conditions in real time, resulting in load distribution deviations and energy efficiency curves that deviate from the optimal range. In addition, existing power grid analyzers and UPS monitoring systems often use discrete sampling and threshold judgment, which cannot capture the coupling relationship between grid power quality (such as harmonics, flicker, and frequency offset) and UPS node characteristics, causing the cluster system to respond lags or even malfunction when the grid is abnormal. Summary of the Invention

[0004] In view of the above problems in the prior art, the present invention provides a UPS energy storage system and method based on power management to solve the above technical problems.

[0005] In order to achieve the above-mentioned and other purposes, the technical solutions adopted by the present invention are as follows:

[0006] In one aspect, the present invention provides a UPS energy storage method based on power management, the method comprising the following steps:

[0007] Step S1: The electrical state parameters of each UPS node in the cluster are collected in real time through the parallel monitoring module, the state parameters of the power grid are collected through the power grid analyzer, the electrical state parameters are subjected to singular value filtering, and a denoised UPS node characteristic data set is generated; a three-dimensional power quality characteristic matrix is ​​constructed based on the power grid state parameters, and an associated mapping is performed with the UPS node characteristic data set;

[0008] Step S2: Perform load characteristic fluctuation phase analysis on the denoised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; predict the load demand range of each UPS node in the next 15 minutes through the LSTM time series network; generate a dynamic energy efficiency map of each UPS node based on historical energy efficiency curve data, and calibrate the optimal energy efficiency working range of each UPS node in combination with the real-time SOC; generate a cluster-level load distribution feasibility matrix based on the cross-comparison between the load demand range and the optimal energy efficiency working range;

[0009] Step S3: Generate a cluster load distribution control instruction set based on the load distribution feasibility matrix, send the load distribution control instruction set to the FPGA controller of each UPS node, and drive the solid-state relay to perform load switching operations.

[0010] Exemplarily, the electrical status parameters of each UPS node include instantaneous input and output voltage values, initial SOC values, and historical energy efficiency curve data; the grid-side status parameters include voltage fluctuation rate, fundamental frequency deviation, and harmonic content characteristic data.

[0011] Exemplarily, step S1 includes:

[0012] Step S11: Based on the synchronous clock, the parallel monitoring module is triggered to perform differential sampling on the instantaneous input and output voltage values ​​of each UPS node to obtain voltage fluctuation time series data; the charging and discharging current of the battery pack is collected in real time through the Hall current sensor, the current data is time-integrated and combined with the initial SOC value to calculate the real-time SOC; the SOC estimate reported by the BMS is simultaneously recorded, and when the deviation between the integral calculated value and the BMS reported value exceeds 3%, the linear compensation algorithm is triggered to generate calibrated SOC data; the historical energy efficiency curve is recorded through a multi-channel DA converter;

[0013] Step S12: extract the fundamental component of the grid voltage through a harmonic analyzer and calculate the periodic deviation rate between the current fundamental frequency and the rated value; use a sliding window algorithm to calculate the voltage peak-to-valley fluctuation amplitude every 10 seconds to generate a voltage fluctuation rate index;

[0014] Step S13: performing singular value decomposition on the voltage fluctuation time series data and the charge and discharge current data to separate high-frequency noise components and cut off interference signals with singular values ​​less than a threshold, thereby generating a denoised UPS node characteristic data set;

[0015] Step S14: Construct a three-dimensional matrix space with voltage fluctuation rate, fundamental frequency deviation and harmonic content as coordinate axes; map the denoised UPS node characteristic data set to the corresponding cells of the matrix according to the timestamp, and generate a spatiotemporal correlation mapping table of power grid quality-equipment status.

[0016] Exemplarily, step S2 includes:

[0017] Step S21: Perform frequency domain decomposition on the denoised UPS node load time series data based on discrete Fourier transform to separate the low-frequency basic load component and the high-frequency fluctuation component; perform time domain phase alignment on the high-frequency fluctuation component, calculate the absolute value of the load change rate within each 5-second window, and generate a time domain load change gradient sequence;

[0018] Step S22: Input the time domain load change gradient sequence into the LSTM time series network, where the input layer of the LSTM time series network includes the current load rate, gradient mean, and ambient temperature data; use the sliding time window mechanism to roll-forward the upper and lower limits of the load demand of each UPS node in the next 15 minutes, and output the probability distribution of the load demand interval;

[0019] Step S23: Extract the efficiency values ​​of each UPS node at different load rates and real-time SOC from the historical energy efficiency curve data to construct a three-dimensional energy efficiency surface; intercept the corresponding slice of the energy efficiency surface according to the real-time SOC value, and calibrate the load rate range with efficiency values ​​greater than 85% as the optimal energy efficiency working range;

[0020] Step S24: Map the probability distribution of the load demand interval to the optimal energy efficiency working interval, and calculate the area ratio of the overlapping area between the two; when the overlapping ratio exceeds 70%, mark it as a feasible allocation node, otherwise mark it as a node to be optimized; summarize the marking status of all nodes to generate a cluster-level load allocation feasibility matrix.

[0021] Exemplarily, step S21 includes:

[0022] A discrete Fourier transform is performed on the denoised load time series data to convert the time domain load signal into a frequency domain amplitude spectrum and phase spectrum. Based on the preset basic load threshold frequency, the low-frequency components below the threshold in the amplitude spectrum are intercepted and restored to the low-frequency basic load time series data through an inverse Fourier transform. The high-frequency components above the threshold frequency are filtered from the frequency domain amplitude spectrum and combined with the original phase spectrum to reconstruct the high-frequency fluctuation time series signal.

[0023] Cross-correlation analysis is performed on the high-frequency fluctuation time series signals of each UPS node to calculate the delay difference between nodes. The phase starting points of all signals are aligned through a cyclic shift operation to generate a phase-synchronized high-frequency fluctuation time series data set. The phase-aligned high-frequency time series data is divided into 5-second windows, and the linear slope of the load value changing with time is fitted in each window. The absolute value of the slope is taken as the load change rate. The load change rate values ​​of all windows are integrated in chronological order to generate a time-domain load change gradient sequence.

[0024] Exemplarily, step S22 includes:

[0025] The time-domain load change gradient sequence is normalized to its minimum-maximum value, and the current load rate, gradient mean, and ambient temperature data are proportionally compressed to the [0, 1] interval to generate an input feature vector with a unified dimension. Taking the current moment as the benchmark, the time-series gradient sequence is intercepted according to a 15-minute historical length window, and the corresponding load rate mean, gradient statistics, and temperature range data within the window are simultaneously extracted to form a multidimensional training sample set. The multidimensional training samples are input into a pre-trained LSTM time series network. The network output layer adopts a bidirectional structure. The first branch predicts the upper limit of future load demand, and the second branch predicts the lower limit. The predicted values ​​are mapped back to the actual dimensional range using a Sigmoid function. Gaussian kernel density estimation is performed on the upper and lower limit results of 50 consecutive rolling forecasts to calculate the probability of the load demand value falling into each interval, and a load demand interval probability distribution table with confidence levels is generated.

[0026] When the difference between the upper and lower limits of the prediction interval exceeds twice the historical maximum fluctuation range, the weighted moving average algorithm is triggered to smooth the prediction results and output the calibrated load demand interval probability distribution.

[0027] Exemplarily, step S23 includes:

[0028] The historical operating data of each UPS node is binned by load factor, divided into 10 intervals ranging from 10% to 100% in 10% steps. The corresponding real-time SOC values ​​within each load factor interval are synchronously correlated, and the average efficiency value for each load factor and SOC combination is extracted to generate a three-dimensional discrete data point set.

[0029] The three-dimensional discrete data point set is input into the thin plate spline interpolation algorithm to generate a continuous and smooth three-dimensional energy efficiency surface with the load rate as the X-axis, the SOC value as the Y-axis, and the efficiency value as the Z-axis. The interpolated surface is then Gaussian filtered to eliminate local noise interference.

[0030] According to the current real-time SOC value, the three-dimensional energy efficiency surface is vertically cut along the Y-axis direction to extract the two-dimensional efficiency-load rate relationship curve that matches the real-time SOC;

[0031] In the two-dimensional efficiency-load rate curve, all load rate points are traversed and continuous segments with efficiency values ​​exceeding 85% are marked. After merging adjacent segments, the optimal energy efficiency working range is generated.

[0032] Exemplarily, step S3 includes:

[0033] Traverse all UPS nodes marked as "feasible allocation nodes" in the load allocation feasibility matrix and generate independent control instruction entries for each node; the instruction entries include target load rate, transfer time window and target energy efficiency threshold parameters;

[0034] According to the priority sequence of nodes in the feasibility matrix, load transfer requests from multiple instruction entries in the same time period are arbitrated; if multiple nodes request to switch loads in the same time window, the node request with a high priority and an energy efficiency value exceeding 85% is responded to first;

[0035] The arbitrated instruction entries are converted into binary control code streams and encapsulated to generate industrial communication data packets. The data packets are broadcast to the FPGA controllers of all UPS nodes via the optical fiber ring network. The FPGA controller generates a PWM modulation signal based on the decoded load parameters, driving the IGBT module of the solid-state relay to switch the load path.

[0036] Another aspect of the present invention provides a UPS energy storage system based on power management, comprising:

[0037] Node data mapping module: This module collects the electrical status parameters of each UPS node in the cluster in real time through the parallel monitoring module, collects the grid-side status parameters through the grid analyzer, performs singular value filtering on the electrical status parameters, and generates a denoised UPS node characteristic data set; constructs a three-dimensional power quality feature matrix based on the grid-side status parameters, and associates and maps it with the UPS node characteristic data set;

[0038] Load matrix generation module: This module performs load characteristic fluctuation phase analysis on the de-noised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; uses an LSTM time series network to predict the load demand range of each UPS node within the next 15 minutes; generates a dynamic energy efficiency map for each UPS node based on historical energy efficiency curve data, and calibrates the optimal energy efficiency operating range of each UPS node in combination with real-time SOC; and generates a cluster-level load distribution feasibility matrix based on a cross-comparison between the load demand range and the optimal energy efficiency operating range.

[0039] Control instruction set generation module: generates a cluster load distribution control instruction set based on the load distribution feasibility matrix, sends the load distribution control instruction set to the FPGA controller of each UPS node, and drives the solid-state relay to perform load switching operations.

[0040] As described above, the UPS energy storage system and method based on power management provided by the present invention have at least the following beneficial effects:

[0041] The power management-based UPS energy storage system and method provided by the present invention collects the electrical status parameters of UPS nodes through parallel monitoring modules and, combined with data from a power grid analyzer, provides a comprehensive understanding of the cluster's operating status. This real-time performance not only improves the timeliness of the data but also helps quickly identify potential faults and abnormal conditions, reducing the risk of system downtime and ensuring the reliability and stability of the power supply. Denoising the electrical status parameters using singular value filtering can effectively improve data quality and reduce misjudgments and incorrect decisions caused by noise. This lays a solid foundation for subsequent load characteristic fluctuation analysis and load demand forecasting, ensuring the accuracy and reliability of the analysis results. Using an LSTM time series network for load demand forecasting can accurately grasp the future load demand of each UPS node. This forward-looking analysis not only helps the system optimize resource allocation but also improves energy efficiency, reduces energy consumption, and thereby achieves improved economic benefits. At the same time, a dynamic energy efficiency map generated based on historical energy efficiency curves, combined with real-time SOC data, can provide guidance on the optimal energy efficiency operating range for each UPS node, ensuring optimal operation under different load conditions. Finally, by cross-comparing the load demand range with the optimal energy efficiency working range, reasonable load distribution control instructions can be formulated to ensure a smooth transition of each UPS node during load switching, reduce the impact of switching, and improve the overall efficiency and safety of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 It is a schematic diagram of the connection of each step of the method of the present invention.

[0044] Figure 2 Schematic diagram of the connection of various modules of the system of the present invention. DETAILED DESCRIPTION

[0045] The above contents described below in conjunction with the implementation of the present invention are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they shall fall within the scope of protection of the present invention.

[0046] The following professional abbreviations used in this manual have their corresponding English and Chinese full names explained as follows:

[0047] LSTM (Long Short-Term Memory): Long Short-Term Memory Network;

[0048] MAP (Maximum A Posteriori): Maximum a posteriori estimation, specifically a parameter estimation method in probability statistics;

[0049] FPGA (Field-Programmable Gate Array): Field-programmable gate array;

[0050] BMS (Battery Management System): Battery Management System;

[0051] SOC (State of Charge): charging state, indicating the percentage of remaining available battery power;

[0052] In the following contents of the specification, the above abbreviations will be mainly used.

[0053] Example 1: Please refer to Figure 1 As shown, a UPS energy storage method based on power management includes the following steps:

[0054] Step S1: The electrical state parameters of each UPS node in the cluster are collected in real time through the parallel monitoring module, the state parameters of the power grid are collected through the power grid analyzer, the electrical state parameters are subjected to singular value filtering, and a denoised UPS node characteristic data set is generated; a three-dimensional power quality characteristic matrix is ​​constructed based on the power grid state parameters, and an associated mapping is performed with the UPS node characteristic data set;

[0055] The electrical status parameters of each UPS node include the instantaneous value of input and output voltage, initial SOC value and historical energy efficiency curve data; the status parameters of the grid end include voltage fluctuation rate, fundamental frequency deviation and harmonic content characteristic data.

[0056] Step S1 includes:

[0057] Step S11: Based on the synchronous clock, the parallel monitoring module is triggered to perform differential sampling on the instantaneous input and output voltage values ​​of each UPS node to obtain voltage fluctuation time series data; the charging and discharging current of the battery pack is collected in real time through the Hall current sensor, the current data is time-integrated and combined with the initial SOC value to calculate the real-time SOC; the SOC estimate reported by the BMS is simultaneously recorded, and when the deviation between the integral calculated value and the BMS reported value exceeds 3%, the linear compensation algorithm is triggered to generate calibrated SOC data; the historical energy efficiency curve is recorded through a multi-channel DA converter;

[0058] Step S12: extract the fundamental component of the grid voltage through a harmonic analyzer and calculate the periodic deviation rate between the current fundamental frequency and the rated value; use a sliding window algorithm to calculate the voltage peak-to-valley fluctuation amplitude every 10 seconds to generate a voltage fluctuation rate index;

[0059] Step S13: performing singular value decomposition on the voltage fluctuation time series data and the charge and discharge current data to separate high-frequency noise components and cut off interference signals with singular values ​​less than a threshold, thereby generating a denoised UPS node characteristic data set;

[0060] Step S14: Construct a three-dimensional matrix space with voltage fluctuation rate, fundamental frequency deviation and harmonic content as coordinate axes; map the denoised UPS node characteristic data set to the corresponding cells of the matrix according to the timestamp, and generate a spatiotemporal correlation mapping table of power grid quality-equipment status.

[0061] In this embodiment of the present invention, the GPS timing module first sends a synchronous trigger signal to the parallel monitoring module, differentially sampling the instantaneous input and output voltages of each UPS node at a 5-millisecond period to generate voltage fluctuation time series data. Simultaneously, a Hall effect current sensor is used to collect the battery pack's charge and discharge currents. The net charge is calculated by integrating the current value with time. This is then combined with the battery's initial SOC value (obtained through full-capacity discharge calibration) to calculate a real-time SOC estimate. Simultaneously, the system receives an SOC estimate based on the open-circuit voltage method reported by the battery management system (BMS). When the integral calculation result deviates from the BMS report by more than 3%, a piecewise linear compensation algorithm is used to dynamically correct the accumulated charge. The calibration coefficient is adjusted based on the direction and magnitude of the deviation, ultimately generating high-confidence SOC data. The voltage, current, and SOC parameters of all nodes are recorded using a 16-bit multi-channel D / A converter, forming a refined historical energy efficiency curve with a timestamp. For grid-side parameters, the harmonic analyzer extracts the fundamental component through a digital phase-locked loop (DPL) and calculates the deviation ratio between the fundamental period and the rated period as the frequency instability rate. Using a sliding time window mechanism, the maximum and minimum grid voltage values ​​within each 10-second window are counted, and the percentage of the difference relative to the rated voltage is calculated as the voltage fluctuation rate indicator. Singular value decomposition is then performed on the voltage fluctuation time series data and the charge and discharge current data matrix. The signal is reconstructed by retaining the top k principal singular value components with an energy contribution exceeding 95%, filtering out high-frequency noise and transient interference to generate a denoised UPS node characteristic dataset. Finally, a three-dimensional power quality feature space is constructed using voltage fluctuation rate (X-axis), fundamental frequency deviation rate (Y-axis), and total harmonic distortion rate (Z-axis). The grid parameters and device status parameters at each time point in the denoised dataset are mapped to matrix elements according to spatiotemporal consistency rules, forming a multidimensional state mapping table with associated attributes.

[0062] Step S2: Perform load characteristic fluctuation phase analysis on the denoised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; predict the load demand range of each UPS node in the next 15 minutes through the LSTM time series network; generate a dynamic energy efficiency map of each UPS node based on historical energy efficiency curve data, and calibrate the optimal energy efficiency working range of each UPS node in combination with the real-time SOC; generate a cluster-level load distribution feasibility matrix based on the cross-comparison between the load demand range and the optimal energy efficiency working range;

[0063] Step S2 includes:

[0064] Step S21: Perform frequency domain decomposition on the denoised UPS node load time series data based on discrete Fourier transform to separate the low-frequency basic load component and the high-frequency fluctuation component; perform time domain phase alignment on the high-frequency fluctuation component, calculate the absolute value of the load change rate within each 5-second window, and generate a time domain load change gradient sequence;

[0065] Step S22: Input the time domain load change gradient sequence into the LSTM time series network, where the input layer of the LSTM time series network includes the current load rate, gradient mean, and ambient temperature data; use the sliding time window mechanism to roll-forward the upper and lower limits of the load demand of each UPS node in the next 15 minutes, and output the probability distribution of the load demand interval;

[0066] Step S23: Extract the efficiency values ​​of each UPS node at different load rates and real-time SOC from the historical energy efficiency curve data to construct a three-dimensional energy efficiency surface; intercept the corresponding slice of the energy efficiency surface according to the real-time SOC value, and calibrate the load rate range with efficiency values ​​greater than 85% as the optimal energy efficiency working range;

[0067] Step S24: Map the probability distribution of the load demand interval to the optimal energy efficiency working interval, and calculate the area ratio of the overlapping area between the two; when the overlapping ratio exceeds 70%, mark it as a feasible allocation node, otherwise mark it as a node to be optimized; summarize the marking status of all nodes to generate a cluster-level load allocation feasibility matrix.

[0068] In an embodiment of the present invention, a discrete Fourier transform is first performed on the denoised load time series data, and the signal is decomposed into a low-frequency basic load component representing steady-state operation and a high-frequency component reflecting transient fluctuations using a preset 0.1Hz threshold frequency. For the high-frequency component, the delay difference between each node is calculated through cross-correlation analysis, and the phase is aligned using a cyclic shift algorithm to extract the phase-synchronized high-frequency fluctuation signal. Subsequently, a linear fit is performed on the signal with a 5-second time window, and the absolute value of its slope is taken as the time domain load change gradient sequence. The gradient sequence, real-time collected load rate mean, and ambient temperature data are input into a pre-trained LSTM timing network. The network uses a sliding time window mechanism to receive 20 minutes of historical data in a rolling manner. The gating unit memorizes long-term and short-term features and outputs the upper and lower limits of the load demand of each node in the next 15 minutes to form a probability distribution interval. At the same time, the efficiency data of each UPS node in the range of 10%-100% load rate and 20%-95% SOC are extracted from the historical database. Cubic spline interpolation is used to construct a three-dimensional energy efficiency surface with load rate as the X-axis, SOC as the Y-axis, and efficiency as the Z-axis. The surface is vertically intercepted according to the current real-time SOC value to obtain a two-dimensional efficiency-load rate relationship curve. The continuous load interval with an efficiency value exceeding 85% is selected as the optimal working section. Finally, the overlapping area between the probability distribution of the load demand interval and the optimal working section is calculated. When the overlap ratio exceeds 70%, it is determined to be a feasible node, otherwise the dynamic optimization mark is triggered. The feasible status marks of all nodes are integrated to generate a binary-coded load distribution feasibility matrix, in which the row and column indexes of each matrix unit correspond to the load carrying capacity of different UPS nodes and time segments.

[0069] Step S21 includes:

[0070] A discrete Fourier transform is performed on the denoised load time series data to convert the time domain load signal into a frequency domain amplitude spectrum and phase spectrum. Based on the preset basic load threshold frequency, the low-frequency components below the threshold in the amplitude spectrum are intercepted and restored to the low-frequency basic load time series data through an inverse Fourier transform. The high-frequency components above the threshold frequency are filtered from the frequency domain amplitude spectrum and combined with the original phase spectrum to reconstruct the high-frequency fluctuation time series signal.

[0071] Cross-correlation analysis is performed on the high-frequency fluctuation time series signals of each UPS node to calculate the delay difference between nodes. The phase starting points of all signals are aligned through a cyclic shift operation to generate a phase-synchronized high-frequency fluctuation time series data set. The phase-aligned high-frequency time series data is divided into 5-second windows, and the linear slope of the load value changing with time is fitted in each window. The absolute value of the slope is taken as the load change rate. The load change rate values ​​of all windows are integrated in chronological order to generate a time-domain load change gradient sequence.

[0072] In an embodiment of the present invention, a discrete Fourier transform is first performed on the denoised load time series data, decomposing the time domain signal into an amplitude spectrum and a phase spectrum in the frequency domain. The amplitude spectrum represents the energy distribution of each frequency component, and the phase spectrum records the timing alignment information of the frequency components. Based on a pre-set basic load threshold frequency (0.1Hz), low-frequency components are extracted from the amplitude spectrum while retaining the corresponding phase spectrum data. Low-frequency basic load time series data reflecting the steady-state operating characteristics of the equipment is reconstructed through an inverse Fourier transform. Meanwhile, the remaining high-frequency components above the threshold are extracted from the amplitude spectrum and combined with the original phase spectrum to perform frequency domain signal synthesis, restoring the high-frequency time series signal containing the transient load fluctuation characteristics. For the high-frequency signals of distributed UPS nodes, a cross-correlation analysis method is used to calculate the time delay difference between the signals of each node. Using the reference node as the reference source, a cyclic shift operation is performed to align the high-frequency fluctuation signals of all nodes in the time domain with the phase starting point, generating a phase-synchronized cluster-level high-frequency time series data set. The phase-aligned high-frequency data are further divided into 5-second time windows. The least squares method is used to fit the linear trend of the load value over time in each window, and the absolute value of the slope of the fitted straight line is taken as the rate of change value representing the intensity of the load fluctuation. Finally, the calculation results of all windows are integrated according to the time sequence of the original data to form a time domain load change gradient sequence that is strictly synchronized with the operating cycle.

[0073] Step S22 includes:

[0074] The time-domain load change gradient sequence is normalized to its minimum-maximum value, and the current load rate, gradient mean, and ambient temperature data are proportionally compressed to the [0, 1] interval to generate an input feature vector with a unified dimension. Taking the current moment as the benchmark, the time-series gradient sequence is intercepted according to a 15-minute historical length window, and the corresponding load rate mean, gradient statistics, and temperature range data within the window are simultaneously extracted to form a multidimensional training sample set. The multidimensional training samples are input into a pre-trained LSTM time series network. The network output layer adopts a bidirectional structure. The first branch predicts the upper limit of future load demand, and the second branch predicts the lower limit. The predicted values ​​are mapped back to the actual dimensional range using a Sigmoid function. Gaussian kernel density estimation is performed on the upper and lower limit results of 50 consecutive rolling forecasts to calculate the probability of the load demand value falling into each interval, and a load demand interval probability distribution table with confidence levels is generated.

[0075] When the difference between the upper and lower limits of the prediction interval exceeds twice the historical maximum fluctuation range, the weighted moving average algorithm is triggered to smooth the prediction results and output the calibrated load demand interval probability distribution.

[0076] In an embodiment of the present invention, minimum-maximum normalization is first performed on the time-domain load change gradient sequence. The gradient value, current load rate, and ambient temperature data are linearly mapped to the interval [0, 1] to eliminate dimensional differences and generate a standardized input feature vector. Simultaneously, based on the current moment, a 15-minute time-series gradient window of data is retrieved along the time axis. The moving average of the load rate within the window, the standard deviation of the gradient sequence, and the temperature extreme value (the difference between the maximum and minimum values) are extracted as multidimensional auxiliary features. Together with the standardized vector, these features form a multidimensional training sample set. The sample set is input into a pre-trained LSTM time series network. The network adopts a bidirectional output layer structure. The first branch uses a time series memory unit to accumulate historical features and predict the upper limit of future load demand. The second branch combines a gating mechanism to capture fluctuation details and predict the lower limit. The output result is restored to the actual load rate percentage dimension through an inverse transformation of the Sigmoid function. Gaussian kernel density estimation is then performed on the upper and lower bounds of 50 consecutive rolling forecasts, with a bandwidth set to 1 / 5 of the standard deviation of historical load fluctuations. The probability density integral of the load demand value within the preset interval is calculated, generating a probability distribution table with a confidence level of 70%-95%. If the width of the forecast interval at a given point in time exceeds twice the maximum historical fluctuation amplitude for that node, the forecast is considered abnormally divergent, triggering a Hamming window-based weighted moving average algorithm. This algorithm constructs a smoothing window by extending five samples forward and backward from the current point, assigning higher weights to intermediate samples for weighted averaging, and outputs a calibrated probability distribution after removing random noise.

[0077] Step S23 includes:

[0078] The historical operating data of each UPS node is binned by load factor, divided into 10 intervals ranging from 10% to 100% in 10% steps. The corresponding real-time SOC values ​​within each load factor interval are synchronously correlated, and the average efficiency value for each load factor and SOC combination is extracted to generate a three-dimensional discrete data point set.

[0079] The three-dimensional discrete data point set is input into the thin plate spline interpolation algorithm to generate a continuous and smooth three-dimensional energy efficiency surface with the load rate as the X-axis, the SOC value as the Y-axis, and the efficiency value as the Z-axis. The interpolated surface is then Gaussian filtered to eliminate local noise interference.

[0080] According to the current real-time SOC value, the three-dimensional energy efficiency surface is vertically cut along the Y-axis direction to extract the two-dimensional efficiency-load rate relationship curve that matches the real-time SOC;

[0081] In the two-dimensional efficiency-load rate curve, all load rate points are traversed and continuous segments with efficiency values ​​exceeding 85% are marked. After merging adjacent segments, the optimal energy efficiency working range is generated.

[0082] In this embodiment, the historical operating data of each UPS node is first binned by load factor, dividing the load factor range from 0% to 100% into ten uniform intervals with a 10% step size. The corresponding real-time SOC values ​​within each load factor interval are then correlated. The arithmetic mean of the efficiency values ​​for each (load factor, SOC) combination is calculated to generate a three-dimensional discrete data point set. This data point set is then input into a thin plate spline interpolation algorithm to construct a continuous energy distribution surface on the two-dimensional plane defined by the load factor (X-axis) and the SOC (Y-axis). Smooth interpolation is achieved by minimizing the curved energy function of the surface. The generated initial surface is then Gaussian filtered with a standard deviation of 0.3, sliding a 3×3 filter window along the XY plane to remove local abrupt noise points. The three-dimensional surface is then sliced ​​perpendicularly along the Y-axis based on the current real-time SOC value to obtain a corresponding two-dimensional efficiency-load factor curve, with the horizontal axis representing the load factor range from 10% to 100% and the vertical axis representing the normalized efficiency value. A sliding window mechanism is used to traverse all load rate points on the two-dimensional curve. When the efficiency values ​​of three consecutive load rate points exceed 85%, they are marked as the starting point of the high-efficiency section. The section is continuously extended until the efficiency value is lower than the threshold, and the sections with adjacent intervals less than 5% load rate are merged. Finally, a set of optimal energy-efficient working intervals containing multiple sub-intervals is generated.

[0083] Step S3: Generate a cluster load distribution control instruction set based on the load distribution feasibility matrix, send the load distribution control instruction set to the FPGA controller of each UPS node, and drive the solid-state relay to perform load switching operations.

[0084] Step S3 includes: traversing all UPS nodes marked as "feasible allocation nodes" in the load allocation feasibility matrix, and generating an independent control instruction entry for each node; the instruction entry includes target load rate, transfer time window and target energy efficiency threshold parameters;

[0085] According to the priority sequence of nodes in the feasibility matrix, load transfer requests from multiple instruction entries in the same time period are arbitrated; if multiple nodes request to switch loads in the same time window, the node request with a high priority and an energy efficiency value exceeding 85% is responded to first;

[0086] The arbitrated instruction entries are converted into binary control code streams and encapsulated to generate industrial communication data packets. The data packets are broadcast to the FPGA controllers of all UPS nodes via the optical fiber ring network. The FPGA controller generates a PWM modulation signal based on the decoded load parameters, driving the IGBT module of the solid-state relay to switch the load path.

[0087] In an embodiment of the present invention, all UPS nodes marked as "feasible" in the load distribution feasibility matrix are first traversed, and based on the real-time load rate and energy efficiency MAP data of the node, independent control instruction entries containing the target load rate, transfer time window and energy efficiency threshold are generated, and each entry is attached with a timestamp and a node ID identifier. According to the predefined priority sequence in the matrix (generated by comprehensive energy efficiency value, response speed and equipment health score), the instructions with load transfer conflicts in the same time window are arbitrated, and node requests with a priority ranking of the top 20% and a real-time energy efficiency value exceeding 85% are responded to first. Secondary time windows are dynamically allocated to low-priority nodes through a weight decay algorithm. The arbitrated instruction entries are encoded according to the Modbus-TCP protocol specification to construct a binary code stream containing a start character, a 32-bit node address, a 16-bit load parameter and a CRC-16 checksum, and encapsulated as an industrial communication data packet.

[0088] Example 2: Please refer to Figure 2 As shown, the UPS energy storage system based on power management includes a node data mapping module, a load matrix generation module and a control instruction set generation module;

[0089] The modules are connected via wired and / or wireless connections to achieve data transmission between the modules;

[0090] Node data mapping module: This module collects the electrical status parameters of each UPS node in the cluster in real time through the parallel monitoring module, collects the grid-side status parameters through the grid analyzer, performs singular value filtering on the electrical status parameters, and generates a denoised UPS node characteristic data set; constructs a three-dimensional power quality feature matrix based on the grid-side status parameters, and associates and maps it with the UPS node characteristic data set;

[0091] Load matrix generation module: This module performs load characteristic fluctuation phase analysis on the de-noised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; uses an LSTM time series network to predict the load demand range of each UPS node within the next 15 minutes; generates a dynamic energy efficiency map for each UPS node based on historical energy efficiency curve data, and calibrates the optimal energy efficiency operating range of each UPS node in combination with real-time SOC; and generates a cluster-level load distribution feasibility matrix based on a cross-comparison between the load demand range and the optimal energy efficiency operating range.

[0092] Control instruction set generation module: generates a cluster load distribution control instruction set based on the load distribution feasibility matrix, sends the load distribution control instruction set to the FPGA controller of each UPS node, and drives the solid-state relay to perform load switching operations.

[0093] It should be noted that the intervals and thresholds are set for ease of comparison. The threshold size depends on the amount of sample data and the cardinality set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless numerical calculations. These formulas are derived from software simulations of the most recent real-world conditions using large amounts of data. The preset parameters in these formulas are set by those skilled in the art based on actual conditions.

[0094] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0095] It should be understood that determining B based on A does not mean determining B based solely on A. B can also be determined based on A and / or other information.

[0096] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0097] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A UPS energy storage method based on power management, characterized in that: The following steps are involved: Step S1: The electrical state parameters of each UPS node in the cluster are collected in real time through the parallel monitoring module, the state parameters of the power grid are collected through the power grid analyzer, the electrical state parameters are subjected to singular value filtering, and a denoised UPS node characteristic data set is generated; a three-dimensional power quality characteristic matrix is ​​constructed based on the power grid state parameters, and an associated mapping is performed with the UPS node characteristic data set; Step S2: Perform load characteristic fluctuation phase analysis on the denoised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; predict the load demand range of each UPS node in the next 15 minutes through the LSTM time series network; generate a dynamic energy efficiency map of each UPS node based on historical energy efficiency curve data, and calibrate the optimal energy efficiency working range of each UPS node in combination with the real-time SOC; generate a cluster-level load distribution feasibility matrix based on the cross-comparison between the load demand range and the optimal energy efficiency working range. The specific operation logic is as follows: By performing frequency domain decomposition and LSTM prediction on UPS load data, the probability distribution of load demand intervals is obtained. A three-dimensional energy efficiency surface is constructed based on historical energy efficiency curves, and the optimal load interval with efficiency greater than 85% is calibrated according to the real-time SOC. The demand interval and the optimal interval are mapped and the overlap ratio is calculated. If the overlap ratio exceeds 70%, it is marked as a feasible node, and finally a cluster load distribution feasibility matrix is ​​generated. The load rate of each UPS node's historical operating data is binned and divided into 10 intervals from 10% to 100% in 10% steps. Synchronously associate the corresponding real-time SOC values ​​within each load rate interval, extract the average efficiency value under each load rate and SOC combination, and generate a three-dimensional discrete data point set; The three-dimensional discrete data point set is input into the thin plate spline interpolation algorithm to generate a continuous and smooth three-dimensional energy efficiency surface with the load rate as the X-axis, the SOC value as the Y-axis, and the efficiency value as the Z-axis. The interpolated surface is then Gaussian filtered to eliminate local noise interference. According to the current real-time SOC value, the three-dimensional energy efficiency surface is vertically cut along the Y-axis direction to extract the two-dimensional efficiency-load rate relationship curve that matches the real-time SOC; In the two-dimensional efficiency-load rate curve, all load rate points are traversed and continuous segments with efficiency values ​​exceeding 85% are marked. The optimal energy efficiency operating range is generated by merging adjacent segments. Step S3: Generate a cluster load distribution control instruction set based on the load distribution feasibility matrix, send the load distribution control instruction set to the FPGA controller of each UPS node, and drive the solid-state relay to perform load switching operations.

2. The UPS energy storage method based on power management according to claim 1, characterized in that: The electrical status parameters of each UPS node include the instantaneous value of input and output voltage, initial SOC value and historical energy efficiency curve data; the status parameters of the grid end include voltage fluctuation rate, fundamental frequency deviation and harmonic content characteristic data.

3. The UPS energy storage method based on power management according to claim 2, characterized in that: Step S1 includes: Step S11: Based on the synchronous clock, the parallel monitoring module is triggered to perform differential sampling on the instantaneous input and output voltage values ​​of each UPS node to obtain voltage fluctuation time series data; the charging and discharging current of the battery pack is collected in real time through the Hall current sensor, the current data is time-integrated and combined with the initial SOC value to calculate the real-time SOC; the SOC estimate reported by the BMS is simultaneously recorded, and when the deviation between the integral calculated value and the BMS reported value exceeds 3%, the linear compensation algorithm is triggered to generate calibrated SOC data; the historical energy efficiency curve is recorded through a multi-channel DA converter; Step S12: extract the fundamental component of the grid voltage through a harmonic analyzer and calculate the periodic deviation rate between the current fundamental frequency and the rated value; use a sliding window algorithm to calculate the voltage peak-to-valley fluctuation amplitude every 10 seconds to generate a voltage fluctuation rate index; Step S13: performing singular value decomposition on the voltage fluctuation time series data and the charge and discharge current data to separate high-frequency noise components and cut off interference signals with singular values ​​less than a threshold, thereby generating a denoised UPS node characteristic data set; Step S14: Construct a three-dimensional matrix space with voltage fluctuation rate, fundamental frequency deviation and harmonic content as coordinate axes; map the denoised UPS node characteristic data set to the corresponding cells of the matrix according to the timestamp, and generate a spatiotemporal correlation mapping table of power grid quality-equipment status.

4. The UPS energy storage method based on power management according to claim 1, characterized in that: Step S2 includes: Step S21: Perform frequency domain decomposition on the denoised UPS node load time series data based on discrete Fourier transform to separate the low-frequency basic load component and the high-frequency fluctuation component; perform time domain phase alignment on the high-frequency fluctuation component, calculate the absolute value of the load change rate within each 5-second window, and generate a time domain load change gradient sequence; Step S22: Input the time domain load change gradient sequence into the LSTM time series network, where the input layer of the LSTM time series network includes the current load rate, gradient mean, and ambient temperature data; use the sliding time window mechanism to roll-forward the upper and lower limits of the load demand of each UPS node in the next 15 minutes, and output the probability distribution of the load demand interval; Step S23: Extract the efficiency values ​​of each UPS node at different load rates and real-time SOC from the historical energy efficiency curve data to construct a three-dimensional energy efficiency surface; intercept the corresponding slice of the energy efficiency surface according to the real-time SOC value, and calibrate the load rate range with efficiency values ​​greater than 85% as the optimal energy efficiency working range; Step S24: Map the probability distribution of the load demand interval to the optimal energy efficiency working interval, and calculate the area ratio of the overlapping area between the two; when the overlapping ratio exceeds 70%, mark it as a feasible allocation node, otherwise mark it as a node to be optimized; summarize the marking status of all nodes to generate a cluster-level load allocation feasibility matrix.

5. The UPS energy storage method based on power management according to claim 4, characterized in that: Step S21 includes: A discrete Fourier transform is performed on the denoised load time series data to convert the time domain load signal into a frequency domain amplitude spectrum and phase spectrum. Based on the preset basic load threshold frequency, the low-frequency components below the threshold in the amplitude spectrum are intercepted and restored to the low-frequency basic load time series data through an inverse Fourier transform. The high-frequency components above the threshold frequency are filtered from the frequency domain amplitude spectrum and combined with the original phase spectrum to reconstruct the high-frequency fluctuation time series signal. Cross-correlation analysis is performed on the high-frequency fluctuation time series signals of each UPS node to calculate the delay difference between nodes. The phase starting points of all signals are aligned through a cyclic shift operation to generate a phase-synchronized high-frequency fluctuation time series data set. The phase-aligned high-frequency time series data is divided into 5-second windows, and the linear slope of the load value changing with time is fitted in each window. The absolute value of the slope is taken as the load change rate. The load change rate values ​​of all windows are integrated in chronological order to generate a time-domain load change gradient sequence.

6. The UPS energy storage method based on power management according to claim 5, characterized in that: Step S22 includes: The time-domain load change gradient sequence is normalized to its minimum-maximum value, and the current load rate, gradient mean, and ambient temperature data are proportionally compressed to the [0, 1] interval to generate an input feature vector with a unified dimension. Taking the current moment as the benchmark, the time-series gradient sequence is intercepted according to a 15-minute historical length window, and the corresponding load rate mean, gradient statistics, and temperature range data within the window are simultaneously extracted to form a multidimensional training sample set. The multidimensional training samples are input into a pre-trained LSTM time series network. The network output layer adopts a bidirectional structure. The first branch predicts the upper limit of future load demand, and the second branch predicts the lower limit. The predicted values ​​are mapped back to the actual dimensional range using a Sigmoid function. Gaussian kernel density estimation is performed on the upper and lower limit results of 50 consecutive rolling forecasts to calculate the probability of the load demand value falling into each interval, and a load demand interval probability distribution table with confidence levels is generated. When the difference between the upper and lower limits of the prediction interval exceeds twice the historical maximum fluctuation range, the weighted moving average algorithm is triggered to smooth the prediction results and output the calibrated load demand interval probability distribution.

7. The UPS energy storage method based on power management according to claim 1, characterized in that: Step S3 includes: Traverse all UPS nodes marked as "feasible allocation nodes" in the load allocation feasibility matrix and generate independent control instruction entries for each node; the instruction entries include target load rate, transfer time window and target energy efficiency threshold parameters; According to the priority sequence of nodes in the feasibility matrix, load transfer requests from multiple instruction entries in the same time period are arbitrated; if multiple nodes request to switch loads in the same time window, the node request with a high priority and an energy efficiency value exceeding 85% is responded to first; The arbitrated instruction entries are converted into binary control code streams and encapsulated to generate industrial communication data packets. The data packets are broadcast to the FPGA controllers of all UPS nodes via the optical fiber ring network. The FPGA controller generates a PWM modulation signal based on the decoded load parameters, driving the IGBT module of the solid-state relay to switch the load path.

8. The UPS energy storage system based on power management is characterized by: The method is implemented based on the UPS energy storage method based on power management according to any one of claims 1 to 7, and includes: Node data mapping module: This module collects the electrical status parameters of each UPS node in the cluster in real time through the parallel monitoring module, collects the grid-side status parameters through the grid analyzer, performs singular value filtering on the electrical status parameters, and generates a denoised UPS node characteristic data set; constructs a three-dimensional power quality feature matrix based on the grid-side status parameters, and associates and maps it with the UPS node characteristic data set; Load matrix generation module: This module performs load characteristic fluctuation phase analysis on the de-noised UPS node characteristic data set to extract the time domain load change gradient of each UPS node; uses an LSTM time series network to predict the load demand range of each UPS node within the next 15 minutes; generates a dynamic energy efficiency map for each UPS node based on historical energy efficiency curve data, and calibrates the optimal energy efficiency operating range of each UPS node in combination with real-time SOC; and generates a cluster-level load distribution feasibility matrix based on a cross-comparison between the load demand range and the optimal energy efficiency operating range. Control instruction set generation module: generates a cluster load distribution control instruction set based on the load distribution feasibility matrix, sends the load distribution control instruction set to the FPGA controller of each UPS node, and drives the solid-state relay to perform load switching operations.

Images