Intelligent control method and system for subway UPS power supply

Abstract

The application relates to the field of subway power supply control and discloses an intelligent control method and system for a subway UPS (Uninterruptible Power Supply) power supply, which comprises the following steps: acquiring original subway state data and power system state indexes; carrying out denoising treatment, passenger flow optimization and load classification on the original subway state data to obtain optimized passenger flow and a passenger flow load classification label; training a neural network model according to the passenger flow load classification label and the optimized passenger flow to obtain a dynamic weight parameter set; training a reinforcement learning agent according to the power system state indexes and the dynamic weight parameter set to obtain a preliminary power distribution strategy; and carrying out stability verification and optimization, risk assessment, preliminary correction, preliminary load balance optimization, noise interference assessment, final correction and final load balance optimization on the preliminary power distribution strategy to obtain a final stable output configuration. The method can improve the stability of a subway power supply system under a dynamic scene.

Description

Technical Field

[0001] This invention relates to the field of subway power supply control, and in particular to an intelligent control method and system for subway UPS power supplies. Background Technology

[0002] In modern urban public transportation systems, subway power systems occupy a central position, and their stability and efficiency directly affect the travel safety and experience of millions of passengers. With accelerating urbanization, subway systems face increasingly complex operating environments, making intelligent control technology a key direction for improving the efficiency and reliability of power systems.

[0003] Traditional power control methods often rely on fixed parameters or single-scenario designs, selecting preset operating modes based on different scenarios to adjust the power control scheme. Firstly, identifying different scenarios requires collecting information from multiple sensors. Existing methods often neglect the correlation between different data sources when processing multi-source data, resulting in a lack of comprehensive basis for generating control commands. Simultaneously, the heterogeneity of data such as passenger flow and equipment operating status requires the system to have efficient data integration capabilities, but current technologies often suffer from inefficiency during data fusion due to the lack of a unified framework, failing to produce results in a timely manner. Secondly, dynamic weight allocation is required when generating power control schemes to adapt to the changing scenarios of the subway. For example, during peak hours, a surge in passenger flow on a certain line can cause sudden changes in equipment load, requiring the system to generate adaptive control commands within seconds. However, existing algorithms struggle to quickly generate accurate weights, thus affecting the timeliness and accuracy of power allocation and leading to equipment instability. For the reasons mentioned above, traditional methods are difficult to adapt to the changing passenger flow and equipment status during subway operation. They lack the ability to perceive and adjust the dynamic environment in real time, resulting in energy waste or insufficient power supply. Especially during peak hours or in case of emergencies, the system response is slow, affecting operational efficiency.

[0004] In summary, traditional power control methods lack the ability to respond to dynamic scenarios in real time and process multi-source data, which affects the flexibility and stability of power adjustment and leads to problems in the subway power system that cannot provide a stable power supply. Summary of the Invention

[0005] This invention provides an intelligent control method and system for subway UPS power supplies, so as to improve the real-time performance, stability and efficiency of subway power supply systems under dynamic environments.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent control method for a subway UPS power supply, comprising: Obtain raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status, and platform environmental parameters; The original subway status data is denoised using Kalman filtering to obtain real-time monitoring indicators and observation noise; among which, the real-time monitoring indicators include optimized passenger flow and the platform environmental parameters. The optimized passenger flow is classified by passenger flow load to obtain passenger flow load classification labels; Based on the passenger flow load classification labels and the optimized passenger flow, a preset neural network model is trained to obtain a dynamic weight parameter set; The preset reinforcement learning agent is trained based on the power system state indicators and the set of dynamic weight parameters to obtain a preliminary power allocation strategy. The stability of the preliminary power allocation strategy is verified and optimized to obtain an optimized control command sequence. A risk assessment is performed on the optimized control command sequence, the dynamic weight parameter set is corrected based on the risk assessment results, and preliminary load balancing optimization is performed based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. The optimized power distribution strategy is evaluated for noise interference. Based on the results of the noise interference evaluation, the optimized power distribution strategy is corrected. Finally, the optimized power distribution strategy is optimized for load balancing to obtain the final stable output configuration.

[0007] Preferably, the original subway status data is denoised using Kalman filtering to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and platform environmental parameters, including: The original subway status data is spatiotemporally registered, and the registered data is denoised using Kalman filtering to obtain a denoised dataset and the observation noise. If the device noise status in the denoised dataset exceeds a preset device noise threshold, then process noise compensation is performed on the passenger flow in the denoised dataset to obtain optimized passenger flow. Based on the optimized passenger flow, the metro status data is fused with real-time collected platform environmental parameters, and the operational status is evaluated to obtain the real-time monitoring indicators.

[0008] Preferably, the step of classifying the optimized passenger flow by passenger load to obtain passenger load classification labels includes: The optimized passenger flow was aligned with a time series to obtain a fluctuating dataset; If the data in the fluctuation dataset exceeds a preset fluctuation threshold, the data exceeding the preset fluctuation threshold is determined to be a peak scene, and a first classification label is obtained; The first category label is updated to obtain the passenger flow load category label.

[0009] Preferably, the step of training a preset neural network model based on the passenger flow load classification labels and the optimized passenger flow to obtain a dynamic weight parameter set includes: The passenger flow load classification labels are standardized to obtain a standardized passenger flow label set; The optimized passenger flow is subjected to feature extraction, and the extracted features are subjected to dimensionality reduction processing to obtain a passenger flow dimensionality reduction feature set; The standardized passenger flow label set and the passenger flow dimensionality reduction feature set are used as inputs to train the preset neural network model, output dynamic weight parameters, and obtain a preliminary set of dynamic weight parameters. The initial dynamic weight parameter set is optimized by using posterior estimation to obtain the optimized dynamic weight parameter set.

[0010] Preferably, the step of training a preset reinforcement learning agent based on the power system state indicators and the dynamic weight parameter set to obtain a preliminary power allocation strategy, and then verifying and optimizing the stability of the preliminary power allocation strategy to obtain an optimized control command sequence, includes: Based on the power system status indicators, the power load is classified to obtain power load classification labels; The power load classification labels are standardized to obtain a standardized power label set; Based on the standardized power label set, feature extraction is performed on the power system status indicators, and the extracted features are then subjected to dimensionality reduction processing to obtain a power dimensionality-reduced feature set. If the power dimensionality reduction feature set meets the preset power threshold condition, then the power dimensionality reduction feature set is input into the reinforcement learning agent for training to obtain the preliminary power allocation strategy; The stability of the initial power allocation strategy is verified using a preset voltage threshold. If the result meets the preset stability index, control commands are generated based on the initial power allocation strategy and a preset optimization strategy, and the optimized control command sequence is generated.

[0011] Preferably, the step of performing a risk assessment on the optimized control command sequence, correcting the dynamic weight parameter set based on the risk assessment results, and performing preliminary load balancing optimization based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy includes: Execute the optimized control command sequence, collect feedback data, and obtain the first response data; If the first response data does not meet the preset power stability index, it is determined that there is a risk of insufficient power supply, and a risk assessment result is obtained. Based on the risk assessment results, the dynamic weight parameter set is corrected through regression analysis, and the optimized control command sequence is corrected based on the corrected dynamic weight parameter set to obtain the corrected command sequence. Execute the correction instruction sequence, collect feedback data, and obtain the second response data; Based on the second response data, the preliminary load balancing optimization is performed to obtain the optimized power allocation strategy.

[0012] Preferably, the step of performing noise interference assessment on the optimized power allocation strategy, correcting the optimized power allocation strategy based on the results of the noise interference assessment, and performing final load balancing optimization based on the corrected optimized power allocation strategy to obtain the final stable output configuration includes: By applying the optimized power distribution strategy, feedback data is collected to obtain third response data; The third response data is subjected to noise interference assessment to obtain the noise interference assessment result; Based on the noise interference assessment results, regression analysis is performed on the optimized power allocation strategy to obtain power conversion adjustment parameters, and the optimized power allocation strategy is corrected based on the power conversion adjustment parameters to obtain a corrected power allocation strategy. The modified power distribution strategy is applied to obtain the fourth response data through a feedback loop; The final load balancing optimization is performed based on the fourth response data to obtain the final stable output configuration.

[0013] Secondly, the present invention provides an intelligent control system for a subway UPS power supply, comprising: The data acquisition module is used to acquire raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status and platform environmental parameters; The data processing module is used to denoise the original subway status data using Kalman filtering to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and the platform environmental parameters; The load classification module is used to classify the optimized passenger flow by passenger flow load and obtain passenger flow load classification labels. The weight generation module is used to train a preset neural network model based on the passenger flow load classification label and the optimized passenger flow to obtain a dynamic weight parameter set. The instruction generation module is used to train a preset reinforcement learning agent based on the power system state indicators and the set of dynamic weight parameters to obtain a preliminary power allocation strategy, and to verify and optimize the stability of the preliminary power allocation strategy to obtain an optimized control instruction sequence. The configuration optimization module is used to perform risk assessment on the optimized control command sequence, correct the dynamic weight parameter set according to the risk assessment result, and perform preliminary load balancing optimization according to the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. The configuration determination module is used to evaluate the noise interference of the optimized power distribution strategy, correct the optimized power distribution strategy based on the results of the noise interference evaluation, and perform final load balancing optimization based on the corrected optimized power distribution strategy to obtain the final stable output configuration.

[0014] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the intelligent control method for the subway UPS power supply described in any of the above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the intelligent control method for the subway UPS power supply described in any one of the above-mentioned methods.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention first acquires the original subway status data, then performs spatiotemporal registration, uses Kalman filtering to denoise the registered data, and finally optimizes it by combining real-time data. This series of operations can effectively handle anomalies and noise in the original passenger flow data, improve the accuracy of passenger flow data, and solve the problem of inaccurate passenger flow data in traditional methods. This invention improves the accuracy and reliability of passenger flow data, providing a more accurate basis for subsequent power distribution strategy adjustments.

[0017] (2) When processing multi-source data, this invention first performs standardization, then feature extraction and dimensionality reduction, and finally generates a solution through machine learning. This series of operations improves the efficiency of processing multi-source data, reduces the burden of data computation, and solves the problems of chaotic data processing and untimely solution generation in traditional methods. This invention optimizes the data processing flow, provides multiple references for the generation of power allocation strategies, and ensures the timeliness and accuracy of the final solution generation.

[0018] (3) After implementing the adjustment scheme, this invention collects response data through a feedback loop and optimizes the scheme in a timely manner based on the response. This feedback optimization is repeated multiple times to achieve the most suitable effect for the current scenario, solving the problem of insufficient flexibility of traditional methods in dynamic scenarios. This invention enriches the diversity of scheme generation and the timeliness of scheme adjustment, making power distribution more intelligent and improving the stability of the power system. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the intelligent control method for subway UPS power supply provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent control system structure of the subway UPS power supply provided in the second embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 The first embodiment of the present invention provides an intelligent control method for a subway UPS power supply, comprising the following steps: Step S1: Obtain raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status, and platform environmental parameters; Step S2: Use Kalman filtering to denoise the original subway status data to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and the platform environment parameters. Step S3: Classify the optimized passenger flow by passenger load to obtain passenger load classification labels; Step S4: Based on the passenger flow load classification label and the optimized passenger flow, train the preset neural network model to obtain a dynamic weight parameter set; Step S5: Train the preset reinforcement learning agent according to the power system state index and the dynamic weight parameter set to obtain a preliminary power allocation strategy. Verify and optimize the stability of the preliminary power allocation strategy to obtain an optimized control command sequence. Step S6: Perform a risk assessment on the optimized control command sequence, correct the dynamic weight parameter set based on the risk assessment results, and perform preliminary load balancing optimization based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. Step S7: Perform noise interference assessment on the optimized power distribution strategy, correct the optimized power distribution strategy based on the noise interference assessment results, and perform final load balancing optimization based on the corrected optimized power distribution strategy to obtain the final stable output configuration.

[0022] In step S1, it is necessary to obtain the original subway status data and power system status indicators; wherein, the original subway status data includes the original passenger flow, equipment noise status and platform environmental parameters.

[0023] In this embodiment, raw subway status data and power system status indicators can be obtained through a sensor network deployed in the subway platform. For example, in the scenario of monitoring passenger flow and equipment noise in the emergency lighting area of ​​the subway platform, the sensor network collects data in real time using a high-frequency sampling method. The sensors collect passenger flow data 10 times per second, such as the number of people passing through per minute, as well as equipment noise status, such as the vibration frequency or current fluctuation value of lighting fixtures, and platform environmental parameters, such as the brightness of lighting fixtures, generating a raw dataset containing timestamps, passenger flow counts, noise amplitude, and platform environmental parameters. Assuming that the passenger flow is 50 people per minute, the equipment noise value is a voltage fluctuation of 0.8 volts, and the brightness of the lighting fixtures is 800 lux during a certain period, the data may be mixed with environmental noise or sensor errors, requiring noise removal and data optimization. This step collects raw data for subsequent data analysis and scheme adjustment.

[0024] In step S2, the original subway status data is denoised using Kalman filtering to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and the platform environment parameters, including: The original subway status data is spatiotemporally registered, and the registered data is denoised using Kalman filtering to obtain a denoised dataset and the observation noise. If the device noise status in the denoised dataset exceeds a preset device noise threshold, then process noise compensation is performed on the passenger flow in the denoised dataset to obtain optimized passenger flow. Based on the optimized passenger flow, the metro status data is fused with real-time collected platform environmental parameters, and the operational status is evaluated to obtain the real-time monitoring indicators.

[0025] In this embodiment, the original subway status data is first time-stamp aligned and spatial coordinate calibrated. In the real-time monitoring scenario of the emergency lighting area of ​​the subway platform, the multi-dimensional data stream collected by the sensor network may include multi-dimensional information such as passenger flow, ambient light intensity, and equipment operating status. Spatiotemporal registration is required to ensure the consistency of this data in time and space. First, time-stamp alignment arranges the data collected by different devices in a time sequence by unifying the time base of the sensors. For example, sensor A of a subway platform collects a passenger flow of 200 people / minute at 10:00:00, and sensor B collects an illumination intensity of 300 lux at 10:00:01. Through time-stamp alignment, the two sets of data are unified to the reference time point of 10:00:00. Then, spatial coordinate calibration is performed to calibrate the sensor position deviations based on the physical layout of the platform. For example, sensors on the east and west sides of the platform might have different installation locations, leading to data discrepancies. The east sensor might collect a passenger flow of 200 people / minute, while the west sensor might collect 202 people / minute. The average passenger flow can be calculated from the average of all sensors to arrive at 201 people / minute, ensuring the data reflects the true state of the same area. After spatiotemporal registration, registered data is obtained. This registration method ensures the accuracy of the initial dataset in both time and space, providing a reliable foundation for subsequent analysis.

[0026] Secondly, the Kalman filter algorithm is used to denoise the registered state data. The Kalman filter estimates the true state based on historical and real-time observation data through prediction and update steps. In this embodiment, passenger flow and equipment noise are used as system state variables, and the observation noise covariance is pre-set based on the historical accuracy data of the subway sensors. First, a state vector is formed using passenger flow and equipment noise states. The predicted state is obtained using the state transition matrix based on the current state, and the prediction error covariance of the predicted state is calculated. The state transition matrix can be determined based on historical data analysis. Then, the Kalman gain matrix is ​​calculated based on the observation noise covariance and the prediction error covariance. Next, the predicted state is updated based on the measured state, the predicted state, and the Kalman gain matrix to obtain the denoised predicted state. The error covariance of the updated predicted state is then calculated to verify the results. For example, assuming the initial state at a certain moment is "passenger flow 50 people / minute, equipment voltage noise 0.5 volts", the state transition matrix yields the next moment's state as "passenger flow 60 people / minute, equipment voltage noise 0.6 volts". The real-time collected state is "passenger flow 55 people / minute, equipment voltage noise 0.8 volts". The prediction error covariance is calculated and combined with the observation noise covariance to obtain the Kalman gain matrix {{0.7, 0.6}, {0.3, 0.4}}. Using the Kalman gain matrix to calculate the denoised state "passenger flow 58.5 people / minute, equipment voltage noise 0.68 volts" from the predicted and observed states, thus generating a denoised dataset. Subtracting the data from the denoised dataset from the original subway state data yields the observation noise. This process effectively reduces the impact of random interference, ensuring the data is closer to the true state and improving the accuracy of subsequent analysis.

[0027] Then, if the equipment noise level in the denoised dataset exceeds a preset equipment noise threshold, process noise compensation is performed on the denoised dataset. For example, if the equipment noise level at a certain data point in the denoised dataset is 0.8 volts, exceeding the preset equipment noise threshold of 0.7 volts, then this data point is determined to be noise exceeding the standard. It should be noted that the equipment noise threshold is derived from historical data on equipment noise levels and passenger flow data. For example, if passenger flow data shows abnormal fluctuations and 80% of the voltage noise exceeds 0.7 volts, then 0.7 volts is set as the voltage noise threshold. Multiple data points are analyzed, and noise compensation is performed on the denoised dataset based on the duration and magnitude of noise exceeding the standard, combined with preset adjustment rules. It should be noted that the adjustment rules are derived from historical data on equipment noise levels and passenger flow data. For example, if a lighting fixture's noise level exceeds 0.7 volts for 10 consecutive minutes, based on historical passenger flow data under similar conditions, it is estimated that this may cause a ±5% fluctuation in passenger flow. A time-domain smoothing filtering algorithm, such as weighted moving average filtering, can be used to remove relevant errors in the denoised dataset and generate optimized passenger flow data. For example, given a 5-minute passenger flow sequence [102, 98, 105, 91, 99] where the fluctuation value 91 exceeds 5% due to lighting noise, filtering is required. For this sequence, a weight distribution of [1, 2, 3, 2, 1] is set according to the rule that the center value has the highest weight, with decreasing weights for the two sides. The smoothing value is calculated as: (102×1 + 98×2 + 105×3 + 91×2 + 99×1) ÷ (1 + 2 + 3 + 2 + 1) = 99 (rounded to the nearest integer). Using the smoothing value 99 to replace the fluctuation value 91 yields the optimized passenger flow [102, 98, 105, 99, 99]. This process ensures that the passenger flow data is not distorted due to equipment failure, guaranteeing the reliability of the monitoring indicators.

[0028] Finally, the optimized passenger flow and real-time collected platform environmental parameters are fused, and the fused data is used to evaluate the operational status. For example, if the optimized passenger flow shows a passenger flow of 50 people per minute, and the platform environmental parameters show a light intensity of 500 lux and a platform temperature of 25 degrees Celsius, one possible implementation is to score the platform environmental parameters based on the passenger flow to obtain an operational status score. When calculating the operational status score, the optimal environmental parameters corresponding to the current passenger flow are first found based on historical data and a preliminary score is given. For example, if the optimal light intensity is 600 lux when the passenger flow is 50 people / minute, and the current light intensity is 400 lux, the corresponding score is 100. 400 / 600 = 66.7 points. Then, based on historical data and current passenger flow, weights are assigned to each environmental parameter for a comprehensive score. For example, when passenger flow is 50 people / minute, temperature is more important, so a weight of 0.7 is assigned to the temperature score, and a weight of 0.3 is assigned to the light intensity score, resulting in scores of 80 and 60 points respectively, yielding an operational status score of 74 points. The raw data and corresponding operational status scores are combined to obtain real-time monitoring indicators, formatted as "Passenger flow: 50 people / minute; Light intensity: 500 lux; Platform temperature: 25 degrees Celsius; Operational status score: 75 points." If the operational status score is lower than the preset scoring standard, such as 75 points while the standard is 80 points (the optimal operational status score is 100; setting the standard to 80 points indicates the system is operating normally), timely adjustments are needed based on environmental parameters. For example, if light intensity is insufficient, the brightness of lighting equipment needs to be increased. This method, through multi-dimensional data integration, improves the comprehensiveness and accuracy of monitoring, helping to promptly identify potential risks.

[0029] In step S3, the optimized passenger flow is classified by passenger flow load to obtain passenger flow load classification labels, including: The optimized passenger flow was aligned with a time series to obtain a fluctuating dataset; If the data in the fluctuation dataset exceeds a preset fluctuation threshold, the data exceeding the preset fluctuation threshold is determined to be a peak scene, and a first classification label is obtained; The first category label is updated to obtain the passenger flow load category label.

[0030] In this embodiment, the optimized passenger flow is first aligned to a time series. Time series alignment ensures data consistency by unifying the time base. For example, in a real-time monitoring scenario of a subway platform, the optimized passenger flow obtained by sensor A after the above steps is displayed as 180 people / minute at 10:00:00, while the optimized passenger flow obtained by sensor B after the above steps is recorded as 190 people / minute at 10:00:02. Through time series alignment, the data is unified to 10:00:00, generating a standard dataset that reflects the trend of passenger flow changes. This alignment method effectively eliminates time discrepancies and ensures the accuracy of subsequent analysis.

[0031] Secondly, the data in the standard dataset is categorized using threshold detection. For example, assuming a normal passenger flow threshold of ±20 people / minute, if the dataset shows a sudden increase in passenger flow from 180 people / minute to 220 people / minute between 10:00:00 and 10:01:00 (a fluctuation of +40 people / minute), exceeding the threshold, it is considered a peak scenario, and the primary category label "Peak" is generated. It should be noted that the passenger flow threshold setting needs to be combined with the platform's historical passenger flow patterns. For example, on platforms far from the city center, where passenger flow is lower, a passenger flow of ±15 people / minute can be considered a peak scenario. This labeling reflects the potential congestion status of the platform, helping to adjust management strategies in a timely manner.

[0032] Then, a supervised learning algorithm is used to update the first classification label. One possible implementation is to use the k-nearest neighbor algorithm. The k-nearest neighbor algorithm optimizes the classification result by comparing the similarity between the current dataset and historical data. First, the features in the standard dataset are standardized, converted to a distribution with a mean of 0 and a standard deviation of 1, and the same operation is performed on the historical data as the input feature vector, combined with the corresponding classification labels as the labeled training dataset. Second, the Manhattan distance is used to calculate the distance between the current data point and every sample point in the training dataset. Then, the k historical data points with the smallest distance to the current point are selected. These k points are the k nearest neighbors of the current point. Finally, a majority vote is performed on the labels of these k nearest neighbors, and the label that appears most frequently is used as the predicted label for the current data point. Specifically, assuming that the historical data includes multiple peak and off-peak scenarios, the k-nearest neighbor algorithm selects the k=5 nearest neighbor samples based on features such as passenger flow and time period, determines whether the current scenario is closer to a peak or off-peak period, and generates a passenger flow load classification label. For example, if the current dataset is similar to 3 peak scenarios and 2 non-peak scenarios, it will eventually be updated to the "peak" label. This iterative approach can dynamically optimize the classification results and improve the reliability of the labels.

[0033] In step S4, based on the passenger flow load classification labels and the optimized passenger flow, a preset neural network model is trained to obtain a dynamic weight parameter set, including: The passenger flow load classification labels are standardized to obtain a standardized passenger flow label set; The optimized passenger flow is subjected to feature extraction, and the extracted features are subjected to dimensionality reduction processing to obtain a passenger flow dimensionality reduction feature set; The standardized passenger flow label set and the passenger flow dimensionality reduction feature set are used as inputs to train the preset neural network model, output dynamic weight parameters, and obtain a preliminary set of dynamic weight parameters. The initial dynamic weight parameter set is optimized by using posterior estimation to obtain the optimized dynamic weight parameter set.

[0034] In this embodiment, the passenger flow load classification labels are first standardized by converting them into numerical values. For example, the load classification labels can be divided into "peak," "valley," or "stable." A minimum-maximum regularization method is used to map the label values ​​to a uniform range of 0 to 1, such as mapping "peak" to 1, "stable" to 0.5, and "valley" to 0. This numericalization facilitates subsequent algorithm processing, ensures that data from different sources have consistent dimensions, effectively reduces data bias, and improves the accuracy of subsequent feature extraction.

[0035] Secondly, features are extracted from the optimized passenger flow data, and these features are then subjected to dimensionality reduction. Specifically, time-dimensional, spatial-dimensional, and behavioral-dimensional features can be extracted from the optimized passenger flow data. In one embodiment, the time-dimensional features include the hourly passenger flow change rate, the spatial-dimensional features include the passenger flow density in different areas, and the behavioral-dimensional features include the average customer dwell time. For example, suppose that the passenger flow change rate of a subway station platform is 20% from 2 pm to 3 pm on Saturday, the passenger flow density in area A is 50 people / square meter, and the average dwell time is 15 minutes. These data constitute a multidimensional feature vector. By performing dimensionality reduction on the multidimensional feature vector through principal component analysis, these high-dimensional features can be compressed into 2-3 principal components. Specifically, in the process of principal component analysis, firstly, each feature is standardized to eliminate the influence of different feature dimensions and scales; secondly, the covariance matrix is ​​calculated to describe the linear correlation between different features; then, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvectors define the direction of the new coordinate axis (the direction of the principal components), while the eigenvalues ​​quantify the magnitude of the data variance along the corresponding eigenvector direction; next, the eigenvectors corresponding to the most important principal components are selected, and these eigenvectors form the principal component matrix; finally, the standardized original matrix and the principal component matrix are multiplied to obtain the dimensionality-reduced passenger flow feature set. For example, after performing principal component analysis on the feature vectors containing three dimensions of information, the two main components of passenger flow change rate and passenger flow density are retained, while the redundant information of average dwell time is removed. This dimensionality reduction reduces computational complexity while retaining key information.

[0036] Then, the standardized passenger flow label set and the passenger flow dimensionality reduction feature set are input into a preset neural network model for training. In this embodiment, the preset neural network model adopts a fully connected neural network structure, consisting of an input layer, hidden layers, and an output layer. The input layer is responsible for receiving a vector composed of the standardized passenger flow label set and the passenger flow dimensionality reduction feature set; the hidden layer contains three hidden layers, each containing 256 neurons, and the ReLU function is selected as the activation function; the output layer contains neurons equal to the number of regions where power allocation is performed, and the output of each neuron is converted into a probability value through the Softmax activation function, forming a preliminary dynamic weight parameter set. The training process of this model goes through four steps: forward propagation, loss calculation, backpropagation, and parameter update. First, random network parameters are set, the standardized passenger flow label set and the passenger flow dimensionality reduction feature set calculated from historical data are used as input, and the power allocation weight parameters that have been verified to be suitable for the corresponding historical scenarios are used as the target output. Then, the cross-entropy loss function is used to compare the deviation between the predicted value and the true value. Next, the gradient of the loss function with respect to all parameters is calculated in reverse. Finally, an optimizer such as Adam is used to update all parameters based on the gradient, and multiple iterations are performed until the weights converge. This dynamic adjustment can improve the model's adaptability to complex scenarios.

[0037] Finally, the initial dynamic weight parameter set is optimized using posterior estimation. Specifically, historical and real-time customer flow data can be combined, and weights can be adjusted through Bayesian inference. For example, suppose historical data for a shopping mall on Saturdays shows that peak hours are concentrated between 2 PM and 4 PM. The current time is 1 PM on Saturday, and customer flow is continuously increasing, approaching a preset threshold. The Bayesian inference method will determine, based on prior knowledge and current data, that the current scenario is approaching a peak period, prioritizing noise interference during this time period and adjusting the weights accordingly. If noise data (such as sensor false alarms) is found to cause weight shifts, the above process can reduce the weight impact of these abnormal data, ultimately resulting in an optimized weight set. This optimization improves the model's robustness to noisy scenarios, ensuring that the generated dynamic weight parameter set better meets actual business needs.

[0038] In step S5, a preset reinforcement learning agent is trained based on the power system state indicators and the dynamic weight parameter set to obtain a preliminary power allocation strategy. The stability of the preliminary power allocation strategy is then verified and optimized to obtain an optimized control command sequence, including: Based on the power system status indicators, the power load is classified to obtain power load classification labels; The power load classification labels are standardized to obtain a standardized power label set; Based on the standardized power label set, feature extraction is performed on the power system status indicators, and the extracted features are then subjected to dimensionality reduction processing to obtain a power dimensionality-reduced feature set. If the power dimensionality reduction feature set meets the preset power threshold condition, then the power dimensionality reduction feature set is input into the reinforcement learning agent for training to obtain the preliminary power allocation strategy; The stability of the initial power allocation strategy is verified using a preset voltage threshold. If the result meets the preset stability index, control commands are generated based on the initial power allocation strategy and a preset optimization strategy, and the optimized control command sequence is generated.

[0039] In this embodiment, firstly, power load is classified according to power system status indicators. These indicators can be obtained from the sensor data stream of the substation, recording hourly electricity consumption, and then categorized by label. For example, if the electricity consumption at a subway platform during a certain period is 5000 kWh, it can be labeled as "high load," 3000 kWh as "medium load," and 2000 kWh as "low load," resulting in power load classification labels. It should be noted that the label classification criteria can be determined by the distribution of electricity consumption in historical data. For example, analyzing the annual electricity consumption of a certain area of ​​the subway platform, if 33% of the data is below 3000 kWh, it is labeled as "low load," and 67% of the data is below 5000 kWh, then 3000 kWh to 5000 kWh is labeled as "medium load," and the remaining data is labeled as "high load." This label classification method facilitates subsequent data analysis and calculation.

[0040] Secondly, the power load classification labels are standardized. One possible implementation is to use a minimum-maximum regularization method to map the labels to the range of 0 to 1. For example, "high load" is mapped to 1, "medium load" to 0.5, and "low load" to 0, resulting in a standardized power label set. This standardization unifies the data format, facilitating subsequent processing.

[0041] Then, features are extracted from the power system state indicators, and the extracted features are subjected to dimensionality reduction. When extracting multidimensional feature vectors from the power system state indicators, features such as the temporal distribution of electricity consumption and regional load can be obtained through statistical analysis of the original data. For example, analyzing electricity consumption in different time periods yields the temporal distribution characteristics of electricity consumption, and analyzing load changes in different regions yields the regional load characteristics. For instance, data from a subway station shows that the electricity consumption change rate from 8:00 AM to 12:00 PM on weekdays is 30%, and the regional load is concentrated in the subway passenger area. This information, along with the power load label, constitutes a multidimensional feature vector. Dimensionality reduction of the multidimensional feature vector through principal component analysis can compress these high-dimensional features into 2-3 principal components, such as the electricity consumption change rate and regional load density, retaining key information and obtaining a power system dimensionality-reduced feature set.

[0042] Next, if the power dimensionality reduction feature set meets the preset power threshold conditions, the power dimensionality reduction feature set and standardized power labels are input into the preset reinforcement learning agent for training. For example, if the electricity consumption change rate in a certain power dimensionality reduction feature set exceeds 20% and the load density is greater than 100 kW / km², then this power dimensionality reduction feature set is input into the reinforcement learning agent for training. It should be noted that the preset power threshold conditions are derived from the analysis of power system status indicators and corresponding electricity demand in historical data. For example, if the electricity consumption change rate of a certain subway station exceeds 20% in 75% of peak hours throughout the year, then 20% is set as the medium electricity consumption change rate threshold. In this embodiment, the reinforcement learning agent adopts the Deep Q-Network (DQN) algorithm, which consists of an action function, a value function, and a prediction model. The input state is a vector composed of the power dimensionality reduction feature set and standardized power labels. The action is a power allocation strategy, including a vector composed of weights allocated to each region. The reward is the change in the operating status score obtained by the aforementioned steps, such as an increase or decrease in the score. After data input, a power allocation strategy is first generated based on the current state using a behavior function. For example, if the load density of the current scenario increases, the power supply to the corresponding area needs to be increased. Then, a value function is used to determine the reward obtained by executing the current strategy. For example, if the power supply is increased and the power demand of the corresponding area is met, a reward of improved operating status score is obtained. Finally, a prediction model is used to predict what state the environment will become next and what immediate reward it will bring. For example, if the power supply is reduced now, the corresponding area will not be able to meet its power demand, resulting in a negative reward of decreased operating status score.

[0043] When training a reinforcement learning agent, power allocation strategies can be optimized through trial-and-error learning and experience replay. First, the prediction model is initialized, the memory is cleared, and a target prediction model is established. Then, the agent selects corresponding actions based on the current state, such as trying a new power allocation strategy or executing a strategy that yields higher rewards based on historical data. Next, the power allocation strategy is executed, the environment transitions to the next state, and a reward is given, such as an increase or decrease in the running state score. The agent then stores this experience as a memory fragment in the memory. Finally, a small batch of experiences is randomly sampled from the memory. For each experience, the target reward value is calculated to make the prediction model closer to the target prediction model. The prediction model is updated through backpropagation and gradient descent. When the agent's performance converges, i.e., the reward no longer significantly increases, the training process is complete. In practice, a power allocation strategy with higher rewards is selected based on the state. For example, if the agent learns from historical data that prioritizing power allocation to critical equipment during peak hours yields higher rewards, then during peak hours, critical equipment such as subway platform elevators and lighting equipment receives 70% power allocation, while other equipment receives 30%. This method of using reinforcement learning agents to generate preliminary power allocation strategies improves the timeliness and accuracy of power adjustments, making the power system more stable.

[0044] Finally, the initial power distribution strategy is validated and optimized for stability, generating an optimized control command sequence. During stability validation, a voltage threshold check can be used to analyze whether voltage fluctuations are within a safe threshold, such as ±5%. For example, a command sequence requires the voltage to be maintained at 220 volts; validation found fluctuations of ±3%, meeting the stability requirement. It should be noted that the safe threshold can be derived from historical data analysis of voltage fluctuations and corresponding power stability scenarios. Afterward, control commands are generated based on a preset optimization strategy. The optimization strategy is determined according to the power supply priority and power demand of different areas. For example, it may stipulate that power is prioritized for the production workshop during peak hours and evenly distributed during off-peak hours. This command sequence ensures efficient and stable power distribution.

[0045] The methods described above are closely aligned with the power distribution needs of smart grids. Standardized label sets unify data formats, dimensionality-reduced feature sets lower processing complexity, reinforcement learning optimizes distribution strategies, and voltage verification and command generation ensure system stability. These steps collectively construct an efficient power management framework that adapts to dynamically changing grid environments.

[0046] In step S6, a risk assessment is performed on the optimized control command sequence, the dynamic weight parameter set is corrected based on the risk assessment results, and preliminary load balancing optimization is performed based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy.

[0047] Execute the optimized control command sequence, collect feedback data, and obtain the first response data; If the first response data does not meet the preset power stability index, it is determined that there is a risk of insufficient power supply, and a risk assessment result is obtained. Based on the risk assessment results, the dynamic weight parameter set is corrected through regression analysis, and the optimized control command sequence is corrected based on the corrected dynamic weight parameter set to obtain the corrected command sequence. Execute the correction instruction sequence, collect feedback data, and obtain the second response data; Based on the second response data, the preliminary load balancing optimization is performed to obtain the optimized power allocation strategy.

[0048] It should be noted that feedback data is typically acquired from data acquisition modules on edge computing devices deployed in substations to ensure real-time performance and accuracy. The data acquisition modules obtain electricity consumption data per second through high-frequency sampling, preferably recording voltage fluctuations and current changes in milliseconds. The first and second response data include voltage, current, load power, and other key indicators.

[0049] In this embodiment, firstly, after executing the optimized control command sequence, a risk assessment is performed on the first response data, comparing it with preset power stability indicators. For example, if the voltage fluctuation threshold is ±5% and the load power limit is 25 kW, and a voltage fluctuation of ±6% or a load power of 26 kW is detected, then a risk of insufficient power supply is determined, and a risk assessment result is obtained. It should be noted that the power stability indicators are statistically derived from historical data on power supply and equipment stability. For example, if the equipment is stable, and the voltage fluctuation is within ±5% 90% of the time, and the load power is less than 25 kW, then these data are used as the power stability indicators.

[0050] Then, the dynamic weight parameter set is adjusted based on the risk assessment results. In one possible implementation, the weight adjustment range of each area's load can be determined based on historical and real-time data. For example, if the current weight of the subway passenger area is 0.7 and other non-critical areas are 0.3, and the risk assessment results show that there is a risk of insufficient power supply in the subway passenger area, and historical data shows that a weight of 0.8 for the subway passenger area can meet the power supply demand under similar conditions, then the weight of the subway passenger area is adjusted to 0.8, and the weight of non-critical areas is reduced to 0.2. Power resources are then redistributed according to the adjusted weight parameter set, generating a new correction instruction sequence.

[0051] Finally, after executing the correction instruction sequence, preliminary load balancing optimization is performed on the second response data. One possible implementation is to use a priority ranking method for preliminary load balancing optimization, prioritizing power supply to high-load areas. For example, a production workshop, due to its dense equipment and long operating hours, is marked as a high-priority area. After weight adjustment, its power allocation ratio is ensured to always be higher than that of low-priority areas, resulting in an optimized power allocation strategy. This method reduces the risk of power shortages by rationally allocating resources, while improving power utilization efficiency. It should be noted that preliminary load balancing optimization may also incorporate characteristics such as regional load density and electricity consumption change rate to further refine the power allocation strategy. For example, if the load density is 120 kW / km² at a certain time, after weight adjustment, power is prioritized for high-density areas to ensure stable grid operation.

[0052] The above method works in conjunction with feedback loops and optimization algorithms to form a closed-loop control mechanism. This mechanism ensures that power distribution adapts to dynamically changing grid demands through real-time feedback and optimization, significantly improving the adaptability and reliability of the grid.

[0053] In step S7, a noise interference assessment is performed on the optimized power allocation strategy. Based on the results of the noise interference assessment, the optimized power allocation strategy is corrected. Finally, a load balancing optimization is performed based on the corrected optimized power allocation strategy to obtain the final stable output configuration, including: By applying the optimized power distribution strategy, feedback data is collected to obtain third response data; The third response data is subjected to noise interference assessment to obtain the noise interference assessment result; Based on the noise interference assessment results, regression analysis is performed on the optimized power allocation strategy to obtain power conversion adjustment parameters, and the optimized power allocation strategy is corrected based on the power conversion adjustment parameters to obtain a corrected power allocation strategy. The modified power distribution strategy is applied to obtain the fourth response data through a feedback loop; The final load balancing optimization is performed based on the fourth response data to obtain the final stable output configuration.

[0054] It should be noted that the third and fourth response data include voltage fluctuations, current intensity, and power noise levels.

[0055] In this embodiment, firstly, an optimized power distribution strategy is executed, and noise interference is assessed on the third response data. During the noise interference assessment, the noise in the response data can be compared with a preset noise interference threshold. For example, the voltage noise interference threshold is 0.4 mV. If the detected voltage noise level is 0.5 mV, it indicates that the noise interference exceeds the standard, and the noise interference assessment result is obtained. This assessment, by quantifying the noise level, provides a basis for adjusting the parameters of the power conversion module. It should be noted that the noise interference threshold needs to be statistically derived by combining historical noise data with the equipment operating status. For example, if the voltage noise is below 0.4 mV 90% of the time when the equipment is operating stably, then 0.4 mV is used as the voltage noise interference threshold.

[0056] Then, the optimized power distribution strategy is revised based on the noise interference assessment results. First, based on the noise interference assessment results, a linear regression algorithm is used to analyze the correlation between historical noise data and real-time data to calculate power conversion adjustment parameters. These parameters include, but are not limited to, the inverter's PWM modulation depth, output voltage amplitude compensation, or switching frequency fine-tuning. For example, under the current load rate and observed high-frequency noise conditions, reducing the inverter unit's PWM modulation depth from the current 0.98 to 0.95 can effectively reduce the amplitude of high-order harmonics generated by power device switching, thereby suppressing the output waveform noise within the voltage interference threshold. Accordingly, the output power conversion adjustment parameter is set to a modulation depth MI of 0.95. The system then modifies the optimized power distribution strategy based on this parameter, generating new inverter control commands to obtain the revised power distribution strategy.

[0057] Finally, after implementing the revised power allocation strategy, the fourth response data undergoes final load balancing optimization. One possible implementation is to use a priority ranking method to optimize power supply to high-load areas. For example, subway passenger areas, due to their high passenger density and safety requirements, are marked as high-priority areas. After weight adjustment, the power allocation ratio for subway passenger areas is consistently higher than that for low-priority areas, resulting in a stable output configuration. It should be noted that the final load balancing optimization may also incorporate load density characteristics and noise variations to adjust the final power allocation strategy. For example, if the load density is 120 kW / km² at a certain time, weight adjustment prioritizes power allocation to high-density areas, ensuring stable grid operation. For instance, if the voltage noise level in the subway passenger area suddenly increases to 0.45 mV during peak hours, the optimized configuration responds quickly, increasing the power conversion efficiency from 92% to 93%, effectively addressing noise interference. This dynamic adjustment mechanism, through real-time feedback and optimization, significantly improves the adaptability and stability of the grid, providing strong support for the reliable operation of the smart grid.

[0058] In summary, this invention discloses an intelligent control method and system for a subway UPS power supply, which acquires raw subway status data and power system status indicators. The raw subway status data includes raw passenger flow, equipment noise status, and platform environmental parameters. Kalman filtering is used to denoise the raw subway status data, obtaining real-time monitoring indicators and observed noise. The real-time monitoring indicators include optimized passenger flow and the platform environmental parameters. The optimized passenger flow is classified by passenger load to obtain passenger load classification labels. Based on the passenger load classification labels and the optimized passenger flow, a preset neural network model is trained to obtain a dynamic weight parameter set. Based on the power system status indicators and... The dynamic weight parameter set is used to train a preset reinforcement learning agent to obtain a preliminary power allocation strategy. The preliminary power allocation strategy is then subjected to stability verification and optimization to obtain an optimized control command sequence. A risk assessment is performed on the optimized control command sequence, and the dynamic weight parameter set is corrected based on the risk assessment results. Preliminary load balancing optimization is then performed based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. Noise interference is assessed on the optimized power allocation strategy, and the optimized power allocation strategy is corrected based on the noise interference assessment results. Finally, final load balancing optimization is performed based on the corrected optimized power allocation strategy to obtain a final stable output configuration. This invention improves the stability of the subway power system in dynamic scenarios.

[0059] Reference Figure 2 The second embodiment of the present invention provides an intelligent control device for a subway UPS power supply, comprising: The data acquisition module is used to acquire raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status and platform environmental parameters; The data processing module is used to denoise the original subway status data using Kalman filtering to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and the platform environmental parameters; The load classification module is used to classify the optimized passenger flow by passenger flow load and obtain passenger flow load classification labels. The weight generation module is used to train a preset neural network model based on the passenger flow load classification label and the optimized passenger flow to obtain a dynamic weight parameter set. The instruction generation module is used to train a preset reinforcement learning agent based on the power system state indicators and the set of dynamic weight parameters to obtain a preliminary power allocation strategy, and to verify and optimize the stability of the preliminary power allocation strategy to obtain an optimized control instruction sequence. The configuration optimization module is used to perform risk assessment on the optimized control command sequence, correct the dynamic weight parameter set according to the risk assessment result, and perform preliminary load balancing optimization according to the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. The configuration determination module is used to evaluate the noise interference of the optimized power distribution strategy, correct the optimized power distribution strategy based on the results of the noise interference evaluation, and perform final load balancing optimization based on the corrected optimized power distribution strategy to obtain the final stable output configuration.

[0060] It should be noted that the intelligent control device for a subway UPS power supply provided in this embodiment of the invention is used to execute all the process steps of the intelligent control method for a subway UPS power supply in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0061] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a data acquisition program. When the processor executes the computer program, it implements the steps in the above-described embodiments of the intelligent control method for subway UPS power supplies, for example... Figure 1 Step S1 is shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data acquisition module.

[0062] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0063] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0064] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0065] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0066] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0067] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A smart control method for a subway UPS power supply, characterized in that, include: Obtain raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status, and platform environmental parameters; The original subway status data is denoised using Kalman filtering to obtain real-time monitoring indicators and observation noise; among which, the real-time monitoring indicators include optimized passenger flow and the platform environmental parameters. The optimized passenger flow is classified by passenger flow load to obtain passenger flow load classification labels; Based on the passenger flow load classification labels and the optimized passenger flow, a preset neural network model is trained to obtain a dynamic weight parameter set; The preset reinforcement learning agent is trained based on the power system state indicators and the set of dynamic weight parameters to obtain a preliminary power allocation strategy. The stability of the preliminary power allocation strategy is verified and optimized to obtain an optimized control command sequence. A risk assessment is performed on the optimized control command sequence, the dynamic weight parameter set is corrected based on the risk assessment results, and preliminary load balancing optimization is performed based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. The optimized power distribution strategy is evaluated for noise interference. Based on the results of the noise interference evaluation, the optimized power distribution strategy is corrected. Finally, the optimized power distribution strategy is optimized for load balancing to obtain the final stable output configuration.

2. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The Kalman filter is used to denoise the original subway status data to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and platform environmental parameters, including: The original subway status data is spatiotemporally registered, and the registered data is denoised using Kalman filtering to obtain a denoised dataset and the observation noise. If the device noise status in the denoised dataset exceeds a preset device noise threshold, then process noise compensation is performed on the passenger flow in the denoised dataset to obtain optimized passenger flow. Based on the optimized passenger flow, the metro status data is fused with real-time collected platform environmental parameters, and the operational status is evaluated to obtain the real-time monitoring indicators.

3. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The process of classifying the optimized passenger flow by passenger load to obtain passenger load classification labels includes: The optimized passenger flow was aligned with a time series to obtain a fluctuating dataset; If the data in the fluctuation dataset exceeds a preset fluctuation threshold, the data exceeding the preset fluctuation threshold is determined to be a peak scene, and a first classification label is obtained; The first category label is updated to obtain the passenger flow load category label.

4. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The step involves training a preset neural network model based on the passenger flow load classification labels and the optimized passenger flow to obtain a dynamic weight parameter set, including: The passenger flow load classification labels are standardized to obtain a standardized passenger flow label set; The optimized passenger flow is subjected to feature extraction, and the extracted features are subjected to dimensionality reduction processing to obtain a passenger flow dimensionality reduction feature set; The standardized passenger flow label set and the passenger flow dimensionality reduction feature set are used as inputs to train the preset neural network model, output dynamic weight parameters, and obtain a preliminary set of dynamic weight parameters. The initial dynamic weight parameter set is optimized by using posterior estimation to obtain the optimized dynamic weight parameter set.

5. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The process involves training a preset reinforcement learning agent based on the power system state indicators and the dynamic weight parameter set to obtain a preliminary power allocation strategy. The stability of the preliminary power allocation strategy is then verified and optimized to obtain an optimized control command sequence, including: Based on the power system status indicators, the power load is classified to obtain power load classification labels; The power load classification labels are standardized to obtain a standardized power label set; Based on the standardized power label set, feature extraction is performed on the power system status indicators, and the extracted features are then subjected to dimensionality reduction processing to obtain a power dimensionality-reduced feature set. If the power dimensionality reduction feature set meets the preset power threshold condition, then the power dimensionality reduction feature set is input into the reinforcement learning agent for training to obtain the preliminary power allocation strategy; The stability of the initial power allocation strategy is verified using a preset voltage threshold. If the result meets the preset stability index, control commands are generated based on the initial power allocation strategy and a preset optimization strategy, and the optimized control command sequence is generated.

6. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The process of performing a risk assessment on the optimized control command sequence, correcting the dynamic weight parameter set based on the risk assessment results, and performing preliminary load balancing optimization based on the corrected dynamic weight parameter set to obtain an optimized power allocation strategy includes: Execute the optimized control command sequence, collect feedback data, and obtain the first response data; If the first response data does not meet the preset power stability index, it is determined that there is a risk of insufficient power supply, and a risk assessment result is obtained. Based on the risk assessment results, the dynamic weight parameter set is corrected through regression analysis, and the optimized control command sequence is corrected based on the corrected dynamic weight parameter set to obtain the corrected command sequence. Execute the correction instruction sequence, collect feedback data, and obtain the second response data; Based on the second response data, the preliminary load balancing optimization is performed to obtain the optimized power allocation strategy.

7. The intelligent control method for subway UPS power supply according to claim 1, characterized in that, The process includes: evaluating the noise interference of the optimized power allocation strategy; revising the optimized power allocation strategy based on the results of the noise interference evaluation; and performing final load balancing optimization based on the revised optimized power allocation strategy to obtain the final stable output configuration. By applying the optimized power distribution strategy, feedback data is collected to obtain third response data; The third response data is subjected to noise interference assessment to obtain the noise interference assessment result; Based on the noise interference assessment results, regression analysis is performed on the optimized power allocation strategy to obtain power conversion adjustment parameters, and the optimized power allocation strategy is corrected based on the power conversion adjustment parameters to obtain a corrected power allocation strategy. The modified power distribution strategy is applied to obtain the fourth response data through a feedback loop; The final load balancing optimization is performed based on the fourth response data to obtain the final stable output configuration.

8. An intelligent control system for a subway UPS power supply, characterized in that, include: The data acquisition module is used to acquire raw subway status data and power system status indicators; wherein, the raw subway status data includes raw passenger flow, equipment noise status and platform environmental parameters; The data processing module is used to denoise the original subway status data using Kalman filtering to obtain real-time monitoring indicators and observation noise; wherein, the real-time monitoring indicators include optimized passenger flow and the platform environmental parameters; The load classification module is used to classify the optimized passenger flow by passenger flow load and obtain passenger flow load classification labels. The weight generation module is used to train a preset neural network model based on the passenger flow load classification label and the optimized passenger flow to obtain a dynamic weight parameter set. The instruction generation module is used to train a preset reinforcement learning agent based on the power system state indicators and the set of dynamic weight parameters to obtain a preliminary power allocation strategy, and to verify and optimize the stability of the preliminary power allocation strategy to obtain an optimized control instruction sequence. The configuration optimization module is used to perform risk assessment on the optimized control command sequence, correct the dynamic weight parameter set according to the risk assessment result, and perform preliminary load balancing optimization according to the corrected dynamic weight parameter set to obtain an optimized power allocation strategy. The configuration determination module is used to evaluate the noise interference of the optimized power distribution strategy, correct the optimized power distribution strategy based on the results of the noise interference evaluation, and perform final load balancing optimization based on the corrected optimized power distribution strategy to obtain the final stable output configuration.

Images