Liquid cooling heat dissipation control system of charging pile

Abstract

The invention discloses a liquid cooling heat dissipation control system of a charging pile, relates to the technical field of energy storage charging piles, and solves the problems of low-frequency oscillation and energy consumption fluctuation caused by independent closed-loop adjustment of a pump and a fan and fault false alarm and failure alarm caused by fixed threshold diagnosis in the prior art. The collection module is used for synchronously collecting and filtering the cooling liquid inlet and outlet temperature, the power module temperature, the pump current and the fan rotating speed; a first prediction modeling module is adopted to generate thermal load prediction and time delay compensation parameters; constructing a self-adaptive baseline relationship between the pump flow and the heat exchange efficiency of the fan through a second baseline modeling module; performing dynamic anomaly judgment based on the residual sequence through a residual detection module; the cooperative scheduling module is used for solving the control track of the pump and the fan by combining the prediction result, the baseline parameter and the abnormal label on the basis of a model prediction control framework; according to the invention, the coordination of liquid cooling heat dissipation control and the adaptability of fault diagnosis are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage charging pile technology, and more specifically to a liquid cooling heat dissipation control system for a charging pile. Background Technology

[0002] With the rapid popularization of new energy vehicles, high-power DC charging piles are gradually becoming mainstream, generating a large amount of heat during operation. Traditional air cooling methods are no longer sufficient to meet heat dissipation requirements, and liquid cooling technology is widely used in charging piles due to its high-efficiency heat exchange capacity. Existing liquid cooling heat dissipation control systems typically rely on the coordinated regulation of water pumps, fans, and sensors to achieve temperature control of the charging module and charging gun cables, thereby ensuring safety and charging efficiency.

[0003] Existing liquid cooling control systems typically achieve temperature regulation through a combination of liquid cooling pumps, fans, and temperature sensors. For example, CN120348175A discloses a charging pile cooling system and control method, which dynamically adjusts the operating status of the liquid cooling pump and fan by real-time sampling of air temperature, gun temperature, and coolant inlet and outlet temperatures, thereby achieving faster and more accurate temperature control and reducing the operating temperature of high-power DC charging piles. This method shows significant improvements in temperature control accuracy and response speed compared to traditional methods. Another document, CN119730197A, proposes a liquid cooler heat dissipation control method. By predicting the gun wire temperature setpoint based on the charging load and combining control with an outer slow-speed loop and an inner fast-speed loop, the coolant flow rate is adjusted in advance, thus improving the lag problem commonly found in traditional heat dissipation control.

[0004] However, the aforementioned technologies still present problems in practice: While solutions like CN120348175A achieve dynamic adjustment of pumps and fans through multi-sensor signals, their control strategies often employ independent closed-loop regulation for each pump and fan, lacking a higher-level coordination mechanism. Under varying dynamic time delays and multi-module load conditions, the pump and fan may compete for performance, leading to low-frequency oscillations in temperature control. This not only increases energy consumption but may also cause derating and noise issues due to temperature fluctuations. On the other hand, while CN119730197A improves system hysteresis through feedforward and cascade control, its fault diagnosis still relies on fixed threshold judgment methods. For example, abnormal flow or temperature is often determined by preset thresholds and time windows. Such static thresholds are difficult to adapt to changes in coolant characteristics, ambient temperature differences, and sensor drift, easily causing false alarms and missed alarms, increasing maintenance costs, and affecting the stability of charging piles. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a liquid cooling heat dissipation control system for charging piles, aiming to improve the coordination of liquid cooling heat dissipation control and the adaptability of fault diagnosis.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution: A liquid cooling heat dissipation control system for a charging pile includes: The data acquisition module is used to establish a unified time-series data stream based on pump current feedback, fan speed feedback and multi-point temperature sampling data through a time synchronization mechanism, and to generate a monitoring dataset using a weighted interpolation method, which is then output to the first prediction modeling module and the second baseline modeling module. The first predictive modeling module is used to generate a future heat load prediction curve based on the monitoring dataset using a time series modeling method, and to obtain the time delay compensation parameters of pumps and fans through cross-correlation phase processing. The prediction results and time delay parameters are output to the collaborative scheduling module in parallel. The second baseline modeling module is used to identify the relationship between circulating pump current, flow rate, fan speed and heat exchange efficiency online using the recursive least squares algorithm. During the identification process, an adaptive baseline model is constructed through a forgetting factor update mechanism, and the baseline prediction value and identification parameter set are output to the residual detection module and the collaborative scheduling module. The residual detection module is used to perform change point analysis based on the baseline predicted values ​​and the monitoring dataset, using a combined method of cumulative sum statistical test and generalized likelihood ratio detection, and output residual sequences, anomaly labels and confidence indices to the collaborative scheduling module. The collaborative scheduling module is used to construct a constrained quadratic programming optimization problem based on the outputs of the first prediction modeling module, the second baseline modeling module, and the residual detection module, and to obtain the joint control trajectory of the circulating pump and the fan by optimizing the solution using the interior point method. The joint control trajectory is output to the drive layer as the final execution command.

[0007] Based on the above technical solutions, the positive and beneficial effects of the present invention are as follows: This scheme inputs the heat load prediction and time delay parameters output by the first predictive modeling module, along with the baseline mapping relationship output by the second baseline modeling module, into the scheduling module for joint optimization within the model predictive control framework. As a result, the control trajectories of the pumps and fans are no longer based on the instantaneous error signals of individual loops, but rather on a constraint optimization solution based on global heat load prediction and cross-correlation time delay compensation. Therefore, the system control quantities tend to be coordinated in the time domain, suppressing the "competition for credit" phenomenon under independent regulation and avoiding low-frequency oscillations and energy consumption fluctuations.

[0008] This solution establishes an adaptive baseline model of pump current-flow rate and fan speed-heat transfer efficiency that updates over time through a second baseline modeling module. In the residual detection module, multi-scale statistical change point analysis of the residual sequence is employed, combined with CUSUM and generalized likelihood ratio methods to form a dynamic discrimination mechanism. Because the diagnostic criteria have shifted from static thresholds to residual discrimination based on the adaptive model, the accuracy of anomaly identification can be improved and the false alarm and false negative rates reduced under varying coolant characteristics and environmental boundary conditions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is an architecture diagram of a liquid cooling heat dissipation control system for a charging pile according to the present invention. Figure 2 This is a schematic diagram illustrating the steps of the cross-correlation phase processing method of the present invention. Figure 3 This is a framework diagram of the first predictive modeling module of the present invention; Figure 4 This is a framework diagram of the adaptive baseline model of the present invention; Figure 5 This is a schematic diagram of the process principle of the collaborative scheduling module of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] As one possible implementation of the present invention: the charging pile adopts a liquid cooling structure near the charging module, with the heat source mainly being the power electronic module and the rectifier / inverter unit. The liquid cooling control system is composed of a data acquisition module, a first predictive modeling module, a second baseline modeling module, a residual detection module, and a collaborative scheduling module, which execute control commands for the circulating pump and fan through the drive layer. The system hardware consists of a sensor subsystem, a drive and execution subsystem, a computing control unit, and a communication and power supply subsystem. The sensor subsystem includes multiple temperature sensors (at least one each at the coolant inlet, coolant outlet, near the power module, and near the heat exchanger; the temperature sensors use industrial-grade platinum resistance thermometers PT100 or high-precision NTC thermistors, with a measurement accuracy better than ±0.2℃); flow / pump current detection uses an integrated Hall current sensor or a shunt in conjunction with an isolated ADC for acquisition, with a recommended resolution of 16 bits and a sampling rate set according to the control loop requirements; the fan speed is acquired by a built-in Hall or hollow shaft speed measurement signal, with an accuracy target of ±1% of the speed. All sensors and drives are connected via industrial fieldbus or Ethernet, with time synchronization using both IEEE 1588 PTP or network time protocol and local hardware timestamps to ensure that the alignment error of multi-channel data on the time axis is controlled at the millisecond level.

[0012] The circulating pump and heat exchanger are positioned as close as possible to the power module being cooled to reduce the liquid loop volume and thermal hysteresis. The heat exchanger (heat sink and fan) is located at the front air intake side of the chassis and uses an air guide structure. The fan is located behind the heat exchanger to create forced convection. The coolant reservoir and gas-liquid separator are placed at the bottom of the chassis to facilitate the removal of air bubbles. The drive layer (inverter, drive board) and control computing unit are placed in an electromagnetically shielded room. The control unit is close to the power module to shorten the sensing and execution loop. All sensor wiring uses shielded twisted-pair cable with grounding. Analog signals pass through an isolated ADC or differential amplifier before entering the controller, and logic signals use optical isolation or isolated transceivers to avoid interference.

[0013] During implementation, such as Figure 1As shown, the overall data flow framework of this invention is as follows: the acquisition module is responsible for real-time data acquisition, frame integrity verification, and time base alignment. During the acquisition process, pump current, fan speed, and multiple temperature points are acquired at different sampling frequencies: pump current and fan speed use higher sampling rates (typically set to 10Hz to 50Hz) to capture rapid dynamic characteristics, while temperature point sampling can be from 1Hz to 5Hz to account for thermal inertia. To ensure the consistency of the sampled data, all sampling frames include timestamps and CRC check fields, and are aligned using PTP time synchronization or a hardware timestamp mechanism. If a sampling frame is lost or exceeds the limit, the acquisition module implements a frame interpolation strategy, using a weighted cubic spline interpolation method to interpolate and align data from different sampling frequencies at a unified time grid. The interpolation weights are set based on the sampling confidence level and the most recent sampling interval. The interpolated output constitutes the monitoring dataset, which is then sent in parallel by the acquisition module to the first predictive modeling module and the second baseline modeling module.

[0014] The first predictive modeling module performs time-series predictions of short-term heat load and cooling demand. Modeling employs ARMA or ARIMA-type models combined with a Kalman smoother to post-process the predicted sequence to reduce noise. The prediction step size is determined by control requirements; in this embodiment, the prediction time domain is 30 to 300 seconds (i.e., short- to medium-term heat load changes), with a time step of 1 or 5 seconds. To compensate for the execution lag of pumps and fans, the prediction module introduces cross-correlation phase analysis during modeling. It determines the typical time lag of the system by calculating the cross-correlation function between pump current and temperature sequences in the monitoring data, and uses this to generate time lag compensation parameters. Cross-correlation phase processing considers changes in environmental conditions; therefore, a sliding window mechanism (e.g., data from the most recent 300 seconds) is used to periodically update the time lag estimate to ensure adaptive adjustment during sudden load changes or system aging. The prediction module outputs the heat load prediction curve and time lag compensation parameters to the collaborative scheduling module in parallel, and provides a prediction confidence estimate for risk weight adjustment in subsequent optimization problems.

[0015] The second baseline modeling module is used to identify the static / quasi-static mapping relationship between circulating pump current-flow rate and fan speed-heat transfer efficiency online. It employs the Recursive Least Squares (RLS) algorithm combined with a forgetting factor to accommodate the slow drift of system parameters over time. Specifically, RLS is used to estimate parameters of several linear or linearized relationships. The forgetting factor ranges from 0.98 to 0.999 and can be adjusted according to operating conditions and sampling frequency to balance response speed and estimation stability. The online identification process receives monitoring datasets from the acquisition module as input, removes transient interference through local window filtering (the filter can be a Kalman filter or a low-pass filter), and triggers re-identification or parameter reset when significant changes in operating conditions are detected (e.g., changes in pump speed or cavitation). The RLS output includes baseline predictions (i.e., the expected flow rate or heat transfer efficiency under the current input) and a set of identified parameters. The baseline predictions serve as a reference for the residual detection module and are also used by the collaborative scheduling module to optimize the linearization of constraints.

[0016] The residual detection module calculates the residual sequence based on baseline predictions and real-time monitoring data and performs change point detection. This embodiment combines the Cumulative Summation Test (CUSUM) with the Generalized Likelihood Ratio (GLR) method: first, the residual sequence is standardized (using a sliding window statistic to estimate the mean and variance), then CUSUM is used to capture small but persistent shifts, while GLR is used to quickly determine single large deviations. The detection module outputs three types of information to the collaborative scheduling module: first, the residual time series for historical backtracking; second, anomaly labels (such as "increased deviation," "sudden jump," "noise surge") indicating the anomaly type; and third, a confidence index, where the confidence level is a floating-point number from 0 to 1 representing the reliability of the detection result. The confidence level is calculated jointly by the residual magnitude, duration, baseline estimation uncertainty, and measurement noise, so that the collaborative scheduling module can implement soft constraints or hard protection for anomalies during optimization.

[0017] The collaborative scheduling module integrates the outputs of the first predictive modeling module, the second baseline modeling module, and the residual detection module into a constrained quadratic programming (QP) problem. In the QP construction, the objective function primarily minimizes temperature deviations and energy consumption at future time points, and introduces penalty terms for the rate of change of control energy and actuator wear to avoid frequent adjustments. The constraint set includes duty cycle boundaries (pump and fan duty cycles between 0% and 100%), rate limits (e.g., pump duty cycle must not change by more than 5% per second, and fan speed must not change by more than 10% per second), minimum start-up and shutdown times, and integral limits to prevent saturation of the control integral term. For abnormal situations (residual detection output confidence exceeding a preset threshold), the collaborative scheduling module can switch relevant constraints to more conservative hard constraints or introduce additional safety margins (such as reducing the maximum duty cycle and increasing fan redundancy) to ensure equipment safety. The QP problem is solved using the interior-point method. The solver is embedded in the edge computing unit and runs in real time. The normal control cycle is set to 1 second, within which the interior-point method converges. If the solution fails, an alternative pre-calculated control trajectory or the most recent safe trajectory is used. The joint control trajectory output by the collaborative scheduling module is the duty cycle and frequency sequence within several future control steps, and the output also includes the execution priority, backoff strategy, and timestamp.

[0018] The driver layer receives control trajectories and converts them into specific PWM or fieldbus commands. The driver layer includes a motor driver, a PWM generator, and an I / O isolation unit. To ensure reliable execution, the driver layer implements a watchdog mechanism, signal integrity verification, and status reporting. The driver gradually adjusts the pump drive current and fan speed according to the trajectory. The driver layer also collects the actual drive response and feeds it back to the acquisition module to form a closed loop. The control system has multiple safety strategies: when the temperature exceeds the upper limit or pump cavitation / stall is detected, the system enters a safety mode, prioritizing temperature control of the cooled components (e.g., temporarily increasing fan speed, operating the pump at a safe point, or triggering intermittent cooling cycles), and, if necessary, executing shutdown or power limiting strategies to protect the entire machine and user safety. Safety events simultaneously trigger alarms and are recorded in local persistent logs and on the remote operation and maintenance platform.

[0019] During implementation, sensor calibration and static characteristic measurements are first performed at the factory or on-site. Temperature, flow rate, and power data are collected under various operating conditions (no load, half load, full load). These data are used to initialize the RLS identification parameters and the initial coefficients of the ARMA model. Online identification is then performed to further converge the parameters. Initial values ​​for control parameters can be based on empirical values; for example, the forgetting factor is set to 0.995, the prediction step size to 120 seconds, the control cycle to 1 second, and the QP rate limit is set to a pump duty cycle variation of no more than 5% / s and a fan duty cycle variation of no more than 10% / s. The integral limit is set with upper and lower limits based on the actuator characteristics. In actual engineering, these parameters can be optimized on-site using successive approximation and gain tuning methods, and safety upper and lower limit configuration options are provided in the software for engineers to adjust.

[0020] To facilitate maintenance and upgrades, the system software is modularly designed. Data acquisition, preprocessing, modeling, change point detection, and optimization run as independent processes or threads, decoupled through an internal message bus. The system supports local log storage and remote reporting (using an encrypted transmission protocol) and provides remote firmware upgrade capabilities. To verify system performance, it is recommended to conduct constant power step tests, random load injection tests, and long-term stability tests on a test bench. The effectiveness of the algorithm can be evaluated by comparing prediction errors, baseline residuals, and energy consumption indicators of control execution. Engineering issues to be considered during testing include sensor noise, the impact of pipeline air bubbles on flow measurement, pump nonlinear characteristics, and minimum operating limits of the fan. These issues should be addressed at the software level using filtering, fault diagnosis, and soft constraints for compensation.

[0021] To facilitate understanding of this embodiment, a detailed description of a liquid cooling heat dissipation control system for a charging pile disclosed in this application embodiment is provided. Please refer to [link to relevant documentation]. Figure 1 The diagram shown is a schematic of the framework of a liquid cooling heat dissipation control system for a charging pile: The data acquisition module is used to establish a unified time-series data stream based on pump current feedback, fan speed feedback, and multi-point temperature sampling data through a time synchronization mechanism, and to generate a monitoring dataset using a weighted interpolation method, which is then output to the first predictive modeling module and the second baseline modeling module. It should be noted that the "pump current feedback" in this application refers to the instantaneous current value acquired by a current sensor (such as a Hall current sensor or a shunt resistor measurement circuit) installed on the power supply side of the liquid-cooled circulating pump. The sampling frequency of this current signal can be set to above 1kHz in implementation to ensure tracking of rapid load fluctuations. The "fan speed feedback" refers to the real-time speed value obtained by a Hall speed sensor, photoelectric encoder, or motor back EMF estimation circuit installed on the fan impeller or motor. The domain of this feedback signal is [0, ...]. ],in This indicates the maximum speed of the fan under rated voltage; "Multi-point temperature sampling data" refers to the temperature values ​​obtained by temperature sensors (such as NTC thermistors, PT100 platinum resistance thermometers, and digital temperature chips) distributed at the inlet and outlet of the liquid cooling circuit pipes, the surface of the cold plate, and the heat dissipation surface of the power devices. The number of temperature sensor locations must be no less than three to ensure that the temperature gradient field of different thermal zones can be covered.

[0022] In a possible implementation, the specific sampling points include: ① a temperature sampling point at the inlet of the liquid cooling loop, located in the pipeline section before the coolant enters the circulating pump from the storage tank or external cooling device; ② a sampling point in the pipeline section after the circulating pump outlet, used to reflect the instantaneous temperature of the coolant after pumping; ③ a sampling point at the cold plate inlet, installed near the heat dissipation surface of the power module; ④ a sampling point at the cold plate outlet, located in the cold plate outlet pipeline, used to measure the temperature after the coolant has carried away heat; ⑤ a sampling point on the heat dissipation surface of the power semiconductor module, typically a thermistor attached to the metal base plate of the IGBT module or MOSFET module; ⑥ auxiliary points, which can be arranged at the inlet and outlet of the coolant storage tank or air-cooled radiator (if applicable). The spatial arrangement of each sensor must ensure that it covers at least the beginning, end, and intermediate nodes of the liquid cooling flow path, forming a closed-loop distribution.

[0023] During the establishment of a unified timing sequence, the time synchronization mechanism is based on cross-channel sampling alignment implemented using a hardware clock synchronization bus (such as the IEEE 1588 PTP protocol or CAN bus synchronization clock), excluding low-precision synchronization methods triggered solely by software polling. Its logical condition is: when all signal samples have a unified timestamp field and the time difference between adjacent samples does not exceed 1ms, synchronization is considered successful; if this threshold is exceeded, interpolation compensation or resampling is triggered.

[0024] The weighted interpolation method is a calculation method that uses adjacent valid sampling points on the time axis for numerical compensation. The formula is: in This indicates the value of the pump current, fan speed, or temperature signal at time t. and Defined as the interpolation weight, specifically set as the reciprocal of the time interval between adjacent sampling points, i.e. This ensures that points with closer temporal distances have greater weight in the interpolation result.

[0025] Layered weighted interpolation is used for temperature signal fusion in the spatial dimension. Specifically, sensors are divided into different layers based on their functional attributes at their locations. For example, the first layer includes coolant inlet and outlet temperature sensors, the second layer includes cold plate surface sensors, and the third layer includes power module surface sensors. The sensor temperatures within each layer... The weighted calculation method is as follows: in Represents the sensor weights, defined in the domain [0,1] and satisfying This weight can be determined based on fluid dynamics simulation results or experience; for example, sensors closer to heat sources are given higher weights.

[0026] In the time domain, the "sliding average" is a local average calculated within a fixed-width time window, and the formula is: in This represents the window length, typically ranging from 5 to 20 sampling points, and can be adjusted based on system thermal inertia. It should be noted that the "sliding average" in this application differs from the global averaging method, and can maintain the dynamic nature of short-term trends.

[0027] For the monitoring dataset, this application defines it as: a structured data unit containing instantaneous values ​​of pump current, instantaneous values ​​of fan speed, fused values ​​of multi-point temperature, and their redundancy check fields, within a unified time-series framework, and encapsulated and output using a frame structure. This frame structure includes at least: ① a frame header (containing timestamps and channel numbers), ② a data area (containing each signal value), ③ a check area (CRC16 or CRC32 checksum), and ④ a frame trailer. It should be noted that the "redundancy check field" in this application is limited to a standard check mechanism capable of detecting bit errors or packet loss errors, and does not include simple parity checks.

[0028] In implementation, the acquisition module may include a high-precision analog-to-digital converter (ADC) with a resolution ≥ 12 bits, a digital signal processor (DSP), a field-programmable gate array (FPGA), and a memory buffer circuit. The ADC is used to sample pump current and analog temperature signals; the DSP or FPGA is used to execute interpolation, weighting, and averaging algorithms; the memory buffer circuit is used to retain at least 50ms of historical data to support sliding average calculation. Logical judgment conditions include: ① If the sensor signal is lost for more than 3 consecutive sampling periods, a missing data compensation mode is triggered; ② If redundancy check fails, the data frame is discarded and a resampling request is triggered; ③ If a sensor value exceeds the physical domain (e.g., temperature <−40℃ or >150℃, current exceeding the pump's rated current), it is marked as an invalid value and does not participate in the weighted calculation.

[0029] In the output after data fusion, the acquisition module not only forms a monitoring dataset but also generates auxiliary indicators for anomaly detection, such as the ratio curve of pump current to fan speed and the rate of change of temperature gradient. These indicators can be used as input features in the subsequent predictive modeling and baseline modeling modules. It should be noted that the auxiliary indicators are derived data obtained through logical operations and function transformations.

[0030] In application implementation, the workflow of the acquisition module can be specifically described as follows: After system startup, an internal timer generates a unified sampling trigger signal, and the pump current sensor, fan speed sensor, and temperature sensor simultaneously sample and store the data in a buffer; subsequently, the DSP calls the interpolation function library to compensate for missing points; the FPGA executes a hierarchical weighted and sliding average algorithm to obtain the fused temperature field; the data is organized into monitoring frames with CRC checksums, and finally transmitted to the upper controller via CAN bus or Ethernet. If a checksum error is detected during the process, the frame is immediately discarded and a retransmission is requested; if a sensor signal exceeds the limit, it is marked and a predefined exception code (such as 0xFFFF) is inserted into the data area.

[0031] Please see Figure 3 The first predictive modeling module is used to generate a future heat load prediction curve based on the monitoring dataset using a time series modeling method, and to obtain the time delay compensation parameters of the pump and fan through cross-correlation phase processing. The prediction results and time delay parameters are output to the collaborative scheduling module in parallel. It should be noted that the first predictive modeling module in this application is different from the existing simple data prediction unit. It is not limited to predicting the trend of coolant temperature change, but through a complete time series modeling, parameter identification and filtering correction process, it outputs a prediction curve that reflects the dynamic heat load of the liquid cooling heat dissipation system, and can provide time delay compensation information in combination with the coupling characteristics of the pump and fan, providing effective support for subsequent control optimization.

[0032] In practical applications, the first predictive modeling module first performs sliding window mean detrending processing. "Detrending" means stripping away the slow, long-term trend, making the series more stable in terms of mean and variance, thus meeting the modeling requirements of the Autoregressive Moving Average (ARMA) model. It should be noted that the "sliding window mean detrending" in this application differs from simple differencing in traditional statistics. It not only adjusts the mean of the data but also removes abnormal peak values ​​within the window to reduce the interference of external shock events on model stability. The length of the sliding window can be determined based on the actual operating frequency and system thermal inertia. For example, when the charging pile operates at 1Hz, the window length can be 60 sampling points corresponding to 1 minute, or it can be extended to 300 sampling points depending on the heat capacity of the cold plate; there is no limitation on this.

[0033] Subsequently, the stationary series was modeled based on an autoregressive moving average model. The autoregressive moving average model consists of an autoregressive component and a moving average component. The autoregressive component characterizes the relationship between predicted values ​​and historical observations, while the moving average component characterizes the relationship between predicted values ​​and historical residuals. Model parameters include autoregressive coefficients. With moving average coefficient The specific values ​​are determined using the maximum likelihood estimation method. During the solution process, the optimal parameter set is determined through iterative optimization of the log-likelihood function. The calculation terminates when the iterative gain is less than a preset threshold, ensuring the stability of the calculation results.

[0034] In the prediction phase, the ARMA model is used to predict future times, obtaining short-term heat load forecasts. Considering the noise interference and measurement biases that exist in actual operation, this embodiment further introduces a Kalman smoother for correction. The Kalman smoother combines recursive filtering and forward backtracking to fuse and correct the ARMA output and real-time observation data. The Kalman gain K is calculated from the state covariance and the observation noise covariance, and its value determines the trade-off between the predicted and observed values. The observation matrix H is used to establish the mapping relationship between the predicted state and the observed values. During the correction process, when the difference between the predicted result and the observed input exceeds a certain range, the Kalman smoother updates the gain by adjusting the value of K, thereby obtaining the corrected future heat load forecast curve.

[0035] In terms of parameter solving mechanism, this application adopts the maximum likelihood estimation method. Its principle is to optimize the likelihood function so that, given the observed data, the estimated model parameters can fit the true data distribution to the maximum extent. It should be noted that the "maximum likelihood estimation" in this application differs from the traditional least-squares-based fitting method. It is suitable for scenarios where the error term approximately follows a Gaussian distribution, and can guarantee the consistency and unbiasedness of the estimated values ​​when the sequence length is large. In terms of computational implementation, the BFGS or L-BFGS algorithms in the numerical optimization library can be called to perform iterative solutions. The iteration stopping condition is that the gradient norm is less than... The specifics can be determined based on the actual situation, and there are no restrictions on them.

[0036] The orders p and q of the ARMA model need to be selected using information criteria (such as AIC and BIC). Typically, p ranges from 1 to 10, and q ranges from 0 to 5, but these can be adjusted according to system complexity and are not strictly limited. The observation matrix H of the Kalman smoother can be set as an identity matrix to directly map the observed input to the state prediction, or it can be defined as a sparse matrix in a multi-dimensional state context to capture the thermal coupling effect between different measurement points. Observation input External disturbances, including coolant flow rate, cold plate temperature distribution, and ambient temperature, can be selected according to the application environment.

[0037] Furthermore, to match the dynamic response characteristics of the pump and the fan, this embodiment employs a cross-correlation phase processing method to extract their time delay parameters. Please refer to [link to relevant documentation]. Figure 2Specifically, by performing cross-correlation calculations on the response curves of the circulating pump flow rate and the fan speed, the phase difference is calculated and defined as a time delay compensation parameter. This time delay compensation parameter, along with the heat load prediction curve, is output to the collaborative scheduling module, enabling subsequent optimization solutions to align the control variables and load predictions in the time dimension. It should be noted that in this application, the cross-power spectral density describes the cross-correlation energy distribution of two sequences in the frequency domain, reflecting their phase relationship and coherence at different frequency components; the phase angle difference specifically refers to the phase difference of the complex spectral quantity across the frequency domain, used to estimate the time delay corresponding to the frequency; the time delay compensation parameter refers to mapping the frequency domain phase difference to a time delay value. This is a scalar or timing curve used by the upper-level scheduler. The specific values ​​can be determined based on the actual situation. Items for which no exact values ​​are specified include window length, overlap rate, and confidence threshold; however, recommended engineering values ​​are provided below for on-site calibration reference.

[0038] In the actual processing flow, the original sequence is first preprocessed: each signal is deDCed and normalized. DeDC can be achieved by removing low-frequency drift using a moving average, and the normalization formula is standardized. ,in and These represent the mean and standard deviation within the sliding window, respectively. The window length and sliding step size are set based on the sampling rate and system response speed. Subsequently, a windowed Fast Fourier Transform (FFT) is used to obtain the spectral characteristics: a Hanning window is used as the window function to reduce sidelobe leakage. The FFT calculation gives... , They are respectively and The frequency domain representation on this window is obtained using the Discrete Fourier Transform (DFT / FFT). Frequency Index Where M is the number of FFT points (usually...) (which can be powers of 2). Based on these spectra, the cross-power spectral density is defined as follows: Simultaneously calculate the single-channel self-power spectrum. To enhance feasibility, the Welch segmented averaging method is used to smooth the spectrum (i.e., averaging the spectrum of each segment after dividing the window into segments), or the spectrum is smoothed in the frequency domain. Perform narrowband smoothing.

[0039] The sequence of cross-correlation functions can be obtained from the inverse Fourier transform of the cross-power spectrum, in the form of: Estimating delay directly using phase angle information in the frequency domain is generally more robust to noise than peak search directly in the time domain; therefore, this embodiment suggests using two types of estimation in parallel: one is frequency domain mapping based on phase difference. Secondly, based on the position of the time-domain cross-correlation peak. To obtain a single usable time delay compensation parameter, the mapping results from different frequencies should be fused according to weights, which can depend on the cross-spectral amplitude or coherence index: define a coherence function. ,when Only when the confidence threshold is exceeded (0.6–0.8 recommended) is the phase mapping at that frequency included in the final delay estimate. The final delay can be given by a weighted average: ,in or All of these are optional strategies. This fusion method can utilize high-energy components while eliminating errors caused by low-coherence noise bands.

[0040] Regarding the implementation of sliding time windows and peak search, it is recommended to use overlapping sliding windows (e.g., window length N samples, overlap rate 50%) on the edge controller to balance resolution and update frequency. Peak detection should be performed on the cross-correlation function during each window slide. The peak position can be refined to subsample points using parabolic fitting or quadratic interpolation, thereby obtaining a time-delay estimate of subsampling accuracy. For non-stationary conditions such as sudden changes in photovoltaic output or charging power, the window length should be temporarily shortened and the update frequency increased after detecting external power fluctuations (e.g., charging power changes by more than a preset percentage threshold, such as 10% / 5s) to quickly capture new dynamic phase characteristics. In steady state, a longer window should be restored to improve the estimated spectral resolution and noise immunity. The triggering conditions are defined as follows: a "fast re-estimation" process is triggered when charging power changes abruptly or PV output fluctuates significantly; a "hold / degrade" strategy is triggered when signal coherence is below the baseline threshold for an extended period or a measurement anomaly is detected, using the last reliable delay value while reducing the weight of delay updates.

[0041] To ensure numerical stability and engineering feasibility, several protection and filtering logics are incorporated into the implementation: the phase wrapping is unwrapped to avoid... Misjudgment caused by jumps; the calculated Implement upper and lower bound limits (e.g., 0.01 s to 5 s, adjustable based on the physical characteristics of the pump and fan); use an exponential smoother or forgetting factor on the weighted average result to avoid short-term jitter affecting the upper-level scheduler. In terms of computational complexity, the FFT has a time complexity of O(n log n). When edge unit resources are limited, segmented small FFT or temporal fast correlation (using a fast cross-correlation algorithm other than FFT) can be used as alternatives; the specific implementation should take into account the real-time control period (e.g., a 1-second control period) so that spectrum estimation and delay mapping can be completed within the control period.

[0042] The specific implementation process is as follows: After reading the signal within the sliding window, DC removal and standardization are performed. A window function is applied and an FFT is calculated. A cross-spectrum is constructed and smoothed. Coherence is calculated and frequency points are selected. The phase of each frequency point is calculated and mapped to a delay. Then, the signals are fused according to weights to obtain the current estimate. Finally, filtering is performed and the result is output to the collaborative scheduling module. If the coherence or signal-to-noise ratio does not meet the preset standard, an "estimation failure flag" is set. When the collaborative scheduling module receives this flag, it should adopt a redundancy strategy (e.g., using the previous reliable τ or using the conservative delay upper limit built into the model).

[0043] The output of this method is a dynamically updated sequence of time-delay compensation parameters. When constructing a quadratic programming or MPC prediction model, the cooperative scheduler incorporates this time delay as an explicit delay or pseudo-state variable into the prediction equation to advance or lag control variables, aligning the execution trajectory with the actual thermal response time. Simultaneously, the anomaly labels and confidence levels from the residual detection module are used to weight the reliability of the delay estimate. If the confidence level is low, the safety margin should be increased and drastic control changes should be avoided. This processing unit should also send several diagnostic quantities estimated in this study (such as dominant frequency, average coherence, and peak SNR) back to the system log and remote operation and maintenance platform for offline analysis and long-term parameter calibration.

[0044] In actual operation, the trigger condition for the first predictive modeling module can be set to start the prediction process when the charging power reaches a preset threshold. For example, when the single-gun output power is detected to exceed 100 kilowatts and the coolant outlet temperature rise rate exceeds the threshold, the module automatically starts and executes predictive modeling to calculate the heat load change trend for the next 10 minutes in advance. This trigger condition can be flexibly adjusted according to factors such as the charging station's power level and cooling requirements, and is not limited thereto.

[0045] It should be noted that the heat load prediction curve in this application not only includes the time evolution trend of the temperature difference between the inlet and outlet of the coolant, but also comprehensively considers the impact of charging power and heat dissipation efficiency on the overall heat load, thus enabling a more comprehensive reflection of the operating status of the liquid cooling system.

[0046] In the implementation process, the first prediction modeling module can run on an embedded controller or on the host computer platform of the charging station. Its program flow can be represented as follows: the data acquisition unit collects real-time monitoring data; the preprocessing unit performs sliding window detrending operation; the modeling unit calls the ARMA model and identifies parameters through the maximum likelihood estimation method; the correction unit calls the Kalman smoother to correct noise in the predicted values; the cross-correlation unit calculates the phase difference between the pump and the fan and generates time delay compensation parameters; finally, the predicted curve and compensation parameters are output to the collaborative scheduling module.

[0047] The prediction curve represents the heat release trend of the power module within a certain time window. The coordinated scheduling module adjusts the pump and fan speeds in advance based on this curve, ensuring the coolant temperature remains within the set range and preventing overheating or overcooling due to delayed adjustments. For example, if the prediction shows that the heat load will increase by 10% within the next 300 seconds, the coordinated scheduling module issues an instruction in advance to gradually increase the pump speed from 3000 rpm to 3600 rpm and the fan speed from 1500 rpm to 1800 rpm, maintaining this trend until the prediction curve inverts. If the Kalman-smoothed curve experiences short-term fluctuations within a certain window but the overall trend remains unchanged, the system will not over-respond, thus ensuring a balance between energy consumption and heat dissipation.

[0048] The second baseline modeling module is used to identify the relationship between circulating pump current, flow rate, fan speed and heat exchange efficiency online using the recursive least squares algorithm. During the identification process, an adaptive baseline model is constructed through a forgetting factor update mechanism, and the baseline prediction value and identification parameter set are output to the residual detection module and the collaborative scheduling module. Specifically, when establishing linear or linearized models of input quantities such as circulating pump current, flow rate, and fan speed with heat exchange efficiency, and using the recursive least squares method to identify parameters online, to prevent excessive parameter jumps caused by abnormal observations or transient shocks, which could compromise the stability of the baseline model, a nonlinear limiting function based on parameter increments is employed to limit the update increment components of each parameter item by item. The limiting threshold is determined by a fixed threshold. Factors that adapt to changing operating conditions The limiting curve is jointly determined by the exponential parameter. The mechanism controls its "soft-to-hard" transition characteristics. It maintains identification accuracy when the increment is small, and suppresses parameter out-of-bounds errors when anomalies occur suddenly, ensuring the identification convergence and the feasibility of the control system.

[0049] Let the input regression vector at time k be... The identified parameter vector is The observed output is The incremental derivation of standard RLS includes Kalman gain and parameter updates: in Let be the covariance matrix (symmetric positive definite, size n×n), and λ∈(0,1] be the forgetting factor (used to give higher weight to recent data). Units and typical suggested values ​​will be given below. To introduce amplitude limiting, the processing formula for each parameter component is defined: The above equation is a term-by-term nonlinear compression mapping, when hour, Keep the original update; when When the denominator tends to a large value, the mapping approximates keeping the increment below the dynamic threshold, thus avoiding single-step jumps. The final parameters are updated using constrained incremental updates. In the formula, As a regression vector, it typically contains pump current, measured or estimated flow rate, fan speed and its hysteresis terms or constructed basis functions, which are dimensionless or whose components are normalized to their original units in practice; For the i-th identification parameter, the unit is related to the inverse of its corresponding regression quantity (for example, if the regression component is flow rate, then the parameter unit is efficiency / flow rate, etc.). Let be the parametric covariance matrix, typically initialized as follows: ( For large numbers, such as ) to express the initial uncertainty; λ is the RLS forgetting factor, typically 0.98–0.9995 (the specific value can be adjusted according to the system parameters, drift rate and sampling frequency). and Units and Consistent; For the baseline threshold of the i-th parameter, it is recommended to express it as a percentage of the absolute value of the initial parameter, for example... (Typically set to 0.05–0.3), or can be given directly as the physically acceptable maximum single-step change (units same as θ). is an adaptive relaxation factor, a scalar greater than or equal to 1, used to temporarily relax the limit when operating conditions fluctuate greatly; p is a nonlinear exponent, p>1, and in engineering, p=2 or p=4 is often chosen. As p increases, the mapping approaches hard truncation. Regarding The specific generation can employ residual or observation variance normalization strategies, for example, by setting a baseline residual sequence. Calculate the variance of the sliding window and reference variance Ratio Construction ,in As an adjustment factor (recommended 0.5–5), this construction ensures that when the residual is amplified... Increase to relax the limits, or vice versa.

[0050] The triggering conditions and action flow must be clearly defined in the engineering implementation: when the observed residual or residual variance within the sliding window Exceeding the preset threshold (For example, a sudden change in instantaneous power or amplified measurement noise) the system enters a "grace mode", that is, temporarily increases the power consumption. (For example, the factor can be relaxed to 2–5), allowing for faster parameter response; the residuals gradually decay after recovering to normal levels and stabilizing over several consecutive steps (e.g., 30–300 s). Going back to point 1. From an implementation perspective, in each update step, the unrestricted increment is first computed. Then, according to the above formula, we get And update In covariance Regarding updates, there are two optional strategies: one is to maintain the traditional RLS. Update (using original) Another approach is to use equivalent correction when the increment is significantly compressed: if and If the difference is significant, the effective gain should be recalculated. and use replace renew To maintain parameter-covariance consistency. To ensure numerical stability, we should... Perform symmetry transformation and lower bound truncation (to ensure that the minimum eigenvalue is not lower than a certain positive number, such as...) ), and the observation matrix is ​​close to singular ( Reinitialization or regularization operations are triggered when the value is very small or very large.

[0051] In a possible practical engineering implementation, the edge controller reads data in each sampling cycle. ,calculate Based on the current Evaluate and update Calculated by nonlinear mapping ,renew Updated later (Choose correction or standard update), and The baseline predicted values ​​are sent to the residual detection and collaborative scheduling module. Fixed-point or double-precision floating-point implementations should be considered to ensure numerical accuracy, and the computational overhead should be evaluated (the complexity of RLS matrix operations increases with the parameter dimension n; typically, in scenarios with n of 4–10, it can run in real-time on industrial-grade ARM processors).

[0052] It should be noted that the "amplitude limiting" in this application differs from simple truncation. This method employs a continuously differentiable compression mapping to avoid the negative impact of discontinuous parameter updates on identification stability. The "forgetting factor" in this paper refers to λ in RLS, which aims to assign higher weights to near-end samples. Unlike the commonly understood "sliding window averaging," it maintains online adaptability in continuously drifting environments. Specific parameters can be determined according to actual conditions; those not limited include window length and reference variance. residual threshold The specific numerical engineering constants should be determined on-site through calibration tests.

[0053] In implementation, restricted parameter updates can significantly reduce transient drift of the baseline model, thus preventing false alarms from the residual detection module due to short-term disturbances. Simultaneously, it provides a stable and reliable baseline prediction for the collaborative scheduling module, avoiding aggressive control commands from the scheduler that could lead to frequent actuator switching or mechanical wear when parameters are abnormal. If the limiting is triggered for an extended period, it can trigger maintenance alarms and recommend manual inspection of sensors or pipeline cavitation and other physical causes to prevent the algorithm from masking persistent faults.

[0054] Additionally, please see Figure 4 The adaptive baseline model set in this application can achieve online iterative updates through forgetting factors and residual statistical feedback, thus possessing the ability to continuously adapt in long-term operating scenarios. Furthermore, the operating condition categories mentioned in this application refer to operating states divided according to the charging power level and the heat load intensity of the liquid cooling circuit, including three categories: low load, medium load, and high load, to facilitate matching appropriate weighted reassembly parameters for different heat load scenarios. Specifically, the threshold range or category boundaries can be determined independently based on the actual operating conditions of the charging station; no limitations are imposed on this.

[0055] In practical operation, the system collects circulating pump current signals, flow meter readings, fan speed feedback, and coolant inlet and outlet temperature differences in each sampling period, and classifies the data according to operating conditions by combining ambient temperature and charging current magnitude. After classification, the recursive least squares algorithm updates the weighted reassembly parameters of the current operating condition category. The update process includes two key elements: first, the introduction of a forgetting factor λ, which assigns decreasing weights to historical samples to ensure that recent data plays a dominant role in model adjustment; second, the calculation of residual statistics, which can be in the form of the sum of squared errors between observed and predicted values, or the mean of residuals accumulated in a sliding window manner. The combination of these two forms a joint evaluation index used to determine the rationality of the current update direction of the model.

[0056] To express this more clearly, the above update mechanism can be represented mathematically as follows: in, This represents the parameter vector at sampling time k. This represents the input feature vector (including pump current, fan speed, etc.). This refers to the measured output value (such as the actual heat exchange efficiency), while This is the gain matrix, whose value is dynamically adjusted according to the combination of residual statistics and the forgetting factor. The significance of the forgetting factor λ in this framework is to ensure that the parameters do not shift over a long period due to historical extreme operating conditions, so that the baseline predictions under different load conditions remain effective.

[0057] When the updated parameters continuously deviate beyond a preset threshold range under a certain operating condition, a limiting constraint mechanism is triggered. For example, under high-load conditions, if the parameter mapping between pump current and flow rate deviates from the nominal value by more than ±15% over several consecutive cycles, the system will automatically limit the update range of this parameter to keep it within a safe boundary, thereby preventing the overall model failure due to sensor malfunctions or local extreme operating conditions. Limiting constraints can be implemented using a simple interval projection method, restricting the updated parameter values ​​to a certain range. Within the range.

[0058] In practical applications, the triggering conditions for this adaptive baseline model include three aspects: First, rapid changes in the heat load of the liquid cooling loop, such as a rapid increase in charging current from 200A to 400A within a short period of time; second, sudden changes in the external ambient temperature, such as a sudden increase in ambient temperature exceeding 5°C under direct sunlight in summer; and third, abnormal fluctuations in the coolant inlet temperature, such as a drop or rise exceeding a threshold within minutes due to unstable operation of the cooling unit. In these scenarios, if the parameters are not corrected in real time through the adaptive baseline mechanism, the mapping relationship between pump current-flow rate and fan speed-heat exchange efficiency will be rapidly distorted, causing the residual detection module to frequently trigger false alarms, and even leading to unreasonable pump and fan control trajectories output by the collaborative scheduling module.

[0059] It should be noted that the "weighted reassembly" described in this application differs from a single fixed weight matrix. Its core feature lies in storing multiple sets of parameters indexed by operating condition categories. It can call different sets of parameters according to different load levels and internally form an update logic linked to residual statistics. For example, under low load conditions, the pump current is less sensitive to flow rate, and the relationship between fan speed and heat exchange efficiency is approximately linear. Therefore, the range of variation of the corresponding parameters in the weighted reassembly is relatively narrow. However, under high load conditions, the influence of fan speed on heat exchange efficiency shows a saturation trend. At this time, the weighted reassembly needs to cover a wider nonlinear range.

[0060] Furthermore, the adaptive baseline model's role in the charging pile liquid cooling control system is not only reflected in parameter identification, but also in providing a reliable benchmark for residual detection. Specifically, when the residual detection module calculates the residual based on the baseline prediction and measured monitoring data, if the baseline model is distorted, the residual will be unnecessarily large, significantly reducing anomaly detection accuracy. By introducing adaptive weight reassembly and amplitude limiting constraints under operating condition classification, this embodiment can ensure the stability of the residual's statistical distribution, thereby enabling subsequent CUSUM or GLR algorithms to output anomaly labels with reasonable confidence levels.

[0061] The residual detection module performs change point analysis based on the baseline predicted values ​​and monitoring dataset, using a combined method of cumulative sum statistical test and generalized likelihood ratio detection. It outputs residual sequences, anomaly labels, and confidence indices to the collaborative scheduling module. The residual detection module operates as follows: it constructs a multidimensional residual vector based on the residual sequences corresponding to the circulating pump current, coolant flow rate, and fan speed, and calculates the residual covariance matrix within each detection window. When the eigenvalues ​​of the covariance matrix exceed a preset anomaly sensitivity range, a secondary discrimination procedure is triggered. This procedure compares the drift trend of the cumulative sum statistic with the likelihood gain function of the generalized likelihood ratio. If both satisfy the anomaly threshold condition, an anomaly label is generated and a confidence index is assigned. Simultaneously, while maintaining the anomaly label state within a continuous detection window, the residual detection module classifies the anomaly mode as a persistent operating condition shift and outputs the classification result to the collaborative scheduling module through a parameter feedback interface. It should be noted that the residual detection in this application not only assesses single-point deviations, but also uses statistical tests and likelihood detection methods to identify trend and structural anomalies in the entire operating sequence, enabling the identification of differences between persistent operating condition deviations and transient disturbances. The module first constructs a residual sequence based on the differences between monitoring data such as circulating pump current, coolant flow rate, and fan speed and baseline predicted values. In practice, the residual sequence is obtained by subtracting the baseline output based on a time series prediction model, a physical mechanism model, or a fusion of both from the real-time signals acquired by the sensors. To avoid biases caused by univariate detection, this module organizes the three types of residual signals into a multidimensional residual vector within the detection window and calculates the residual covariance matrix. Elements of the matrix This reflects the relationship between the i-th monitoring quantity and the th... The correlation of residuals among various monitoring quantities, and its eigenvalues. In a statistical sense, it describes the overall level of dispersion of the system residuals.

[0062] In practice, when the largest eigenvalue of the covariance matrix... Exceeding the preset sensitivity range At this point, the system enters the secondary discrimination stage. It should be noted that the sensitivity interval in this application is dynamically calibrated based on historical system operating data to ensure a stable detection rate under different loads, coolant viscosities, or ambient temperatures. The specific calculation method for the upper and lower thresholds can be determined according to actual conditions, such as based on quantile statistics of empirical distribution functions or derivation based on confidence intervals under maximum likelihood estimation; no limitations are imposed on this method.

[0063] The secondary discrimination procedure simultaneously invokes the cumulative sum statistical test (CUSUM) and the generalized likelihood ratio (GLR) test. The core of the CUSUM method lies in constructing the statistical measure. ,in Let be a projection of the residual vector. is the baseline mean, and k is the drift-sensitive parameter. When A sustained upward trend indicates a potential system offset. GLR detection, on the other hand, is based on the likelihood gain function. Where θ represents the set of parameters under the anomaly model, This represents the parameters under the normal model. If... If the value exceeds the preset threshold η, it is considered abnormal. The module only generates an abnormal label when both the CUSUM drift trend and the GLR likelihood gain meet the threshold conditions.

[0064] The specific actions include: once labels are generated, the module encodes the residual statistical features (such as the distribution of mean, variance, kurtosis, and covariance matrix eigenvalues) within the current detection window into feature vectors, which serve as the basis for the confidence index. The confidence score can be calculated using the Sigmoid function. The multidimensional statistical features are normalized, and the output range is limited to [0,1], representing the reliability level of the anomaly detection. This indicator, along with the anomaly label, is sent to the collaborative scheduling module to determine whether to adjust the pump speed, fan speed, or trigger the protection logic of the liquid cooling cycle.

[0065] When an anomaly tag remains true for multiple consecutive detection windows, the module further classifies it as a "persistent operating condition offset." It should be noted that in this application, "persistent operating condition offset" differs from "transient disturbance." The former refers to structural problems such as persistent overload or decreased heat transfer efficiency in the cooling circuit, while the latter is merely a short-term fluctuation. The classification logic relies on a tag duration threshold τ. When a tag persists for more than τ, the system triggers the classification procedure and transmits the result to the collaborative scheduling module through the parameter feedback interface. Here, τ can be determined based on the system's thermal inertia and control response time, and is not limited thereto.

[0066] Eigenvalue decomposition formula of covariance matrix ,in It is a diagonal matrix containing eigenvalues. By monitoring and The ratio of the residual distribution to the mean can determine whether the residual distribution exhibits abnormal stretching or shrinkage. Both CUSUM and GLR algorithms rely on sequence probability modeling; the former is suitable for detecting mean shift, while the latter can adaptively estimate the intensity of anomalies under unknown parameter conditions. Combining the two can reduce the false alarm rate and improve the reliability of anomaly labels.

[0067] In the liquid cooling control system of charging piles, this residual detection module acts as a pre-positioned safety diagnostic unit. When the charging power suddenly increases or the external ambient temperature changes drastically, the residual difference between the coolant flow rate and the pump current often exhibits a coupling shift. Relying solely on a single monitoring quantity may lead to missed detections or false judgments. However, multi-dimensional residual and statistical detection methods can comprehensively determine whether the system truly has hidden dangers such as insufficient heat dissipation or liquid path obstruction. Once the anomaly label is confirmed, the collaborative scheduling module will determine whether to activate the backup cooling fan, increase the pump speed, or reduce the power output of some charging modules based on the confidence index, to ensure that the liquid cooling system remains within safe limits.

[0068] The collaborative scheduling module is used to construct a constrained quadratic programming optimization problem based on the outputs of the first prediction modeling module, the second baseline modeling module, and the residual detection module, and to obtain the joint control trajectory of the circulating pump and the fan by optimizing the solution using the interior point method. The joint control trajectory is output to the drive layer as the final execution command.

[0069] For details, please refer to Figure 5 The coordinated scheduling module executes at the control level, and its inputs include: the heat load prediction curve given by the first predictive modeling module. With time delay compensation The identification parameters provided by the second baseline modeling module The anomaly labels provided by the residual detection module With confidence level The sampling period is denoted as... The prediction time domain length is N steps. It should be noted that in this application, the "duty cycle boundary" limits the pump PWM duty cycle command. closed interval and fan speed control commands closed interval "Rate limit" restricts the variation range between adjacent sampling steps. "Points Limit" restricts points status. interval ,in , , To predict temperature indicators (by (obtained through thermal model mapping) This serves as a temperature reference trajectory. The aforementioned upper and lower bounds and rate thresholds are determined based on the device datasheet, thermal inertia, and pump / fan mechanical limits. Specific values ​​can be provided based on field calibration and safety strategies, and are not limited in this regard.

[0070] Input alignment stage Perform time-series translation: and the aligned Merge into heat demand vector Thermal dynamics are adopted through Linearized discrete model: ,in This is a predicted quantity of temperature or inlet / outlet temperature difference. For the disturbances caused by ambient temperature and inlet liquid temperature, matrices A, B, C, and E are determined by the Jacobian of the cold plate, pump head-flow characteristics, and fan heat exchange efficiency surface at the current operating point. It should be noted that the "cooperative scheduling" in this application is not a single-channel PI superposition, but rather a multivariate trajectory optimization constructed based on the aforementioned coupled state equations.

[0071] The objective function is a quadratic form within the prediction domain: , It is a symmetric positive definite weight matrix. Let be the integral state penalty coefficient. The constraint set is given as follows: To facilitate the solution, the absolute value constraint is rewritten as a linear inequality: This leads to the standard convex quadratic programming problem: , where z stack H is from The model matrix is ​​generated by combination, maintaining symmetry and positive definiteness.

[0072] KKT System Construction: Initial Feasibility Duality Complementary relaxation Equilibrium conditions ,in For equality-bound multipliers, For inequality constraint multipliers, Writing the Lagrange function Interior point method introduces obstacle parameters With logarithmic obstacle term Newton's step is obtained from the augmented system: in Backtracking search selects the step size. The inequality balance and complementarity monotonically decrease; the obstacle parameters are as follows: renew, The termination criteria employ three threshold criteria: original residual, dual residual, and complementary gap. .

[0073] The abnormal coupling strategy is triggered when L=1, and the confidence level c enters the penalty mapping: And add a regularization term that deviates from the baseline. As the baseline trajectory, Rate threshold tightening rule: Duty cycle interval tightening rule The above. Tuning is performed through offline simulation or online testing, with the numerical range consistent with the device safety boundary; no restrictions are imposed on this range. If L=0, the weights revert to their nominal values ​​based on the boundary conditions.

[0074] The instruction sequence frame adopts the "control trajectory frame" format: header field [0xAA55], timestamp (milliseconds), step number k, pump duty cycle. (Fixed point Q15), fan speed (rpm fixed-point Q15), CRC-16 checksum. The transmission link supports CAN2.0B, RS-485, or SPI direct connection. Clock synchronization uses PTP sublayer or local hardware timer alignment, and drift error is limited to... The driver layer receiver performs boundary trimming again on the duty cycle and speed commands. To ensure that the hardware does not exceed the limits; the pump driver uses PWM complementary output, and the dead zone is configured according to the device manual; the fan driver uses V / f or closed-loop speed loop, and the speed measurement input comes from Hall or FG pulse.

[0075] The trigger-solve-deployment cycle is defined as a rolling time-domain strategy: that is, each Perform a solution once; if the residual label jumps from 0 to 1, immediately trigger a weighted reconstruction and re-solve once in the current period; if Or external disturbance estimates Crossing the threshold This triggers a retry during the process. Threshold The settings are based on historical distribution quantiles or operational experience. To avoid instruction jitter, a first-order feedforward filter is added to the output side. When the filter parameter L=1, it is used according to... Tighten, .

[0076] in, Each step, The fan start-up threshold is set to be no less than 0.15 times the rated speed. Set the upper limit of the rating to 1.00 times the rated value. Each step, The values ​​are determined by the linkage between steady-state error requirements and anti-integral saturation strategies. The above values ​​are adjusted according to hardware differences and are not limited thereto.

[0077] During implementation, the predictive modeling output Enter the timing alignment unit to generate Baseline identification provides Provides state equation assembly and constraint tuning; residual detection provides (L,c) driving weight reconstruction and boundary tightening; optimization kernel is generated according to the KKT-internal point method. Trajectory; Execution of the driving layer And feedback measurements are used as the starting point for state estimation and residual calculation in the next cycle. It should be noted that "anomaly label confidence coupling weights" refers to the weights of c on... Monotonic mappings within boundary intervals are used to avoid discrete jumps; "integral limiting" refers to limiting the integral amplitude at the level of optimization variables. By directly adding box constraints instead of controller backend clamping, we can ensure that both the optimality condition and physical feasibility are met simultaneously.

[0078] In implementation, the upper bound of the round-trip delay between the control trajectory frame and the driving layer. Clock synchronization error CRC check failure triggers frame dropping and reuses the previous cycle. When the cumulative frame loss count reaches a threshold M, a protection derating is triggered, where M is set by the system security policy. Boundary clipping is recorded synchronously, with the record items including the clipping channel, amplitude, and timestamp, for use in logging and online tuning.

[0079] In the further refinement of the collaborative scheduling module, the introduction of a penalty function mechanism aims to dynamically adjust the objective weights of the optimization problem when an abnormal state is detected in the liquid cooling heat dissipation control system. This ensures that the system can still obtain a stable joint control trajectory of the circulating pump and fan through adaptive adjustment when subjected to disturbances (such as photovoltaic fluctuations, sudden load changes, cooling circuit blockage, or fan jamming). The "penalty function mechanism" in this application differs from the traditional method of accumulating single constraint penalties. By dynamically generating error terms and switching penalty terms through anomaly intensity coefficients and piecewise linear mapping relationships, it endows the optimizer with the ability to sensitively respond to abnormal states at the mathematical model level.

[0080] The "anomaly label" and "confidence index" output by the residual detection module are used together as trigger inputs. The anomaly label is a binary signal (0 represents normal, 1 represents anomaly), and the confidence index is a real value in the interval [0,1]. An anomaly intensity coefficient γ is defined as follows: γ = label × conf (where label∈{0,1}, conf∈[0,1]), ensuring that the value of γ increases significantly when an anomaly exists and the confidence level is high. The specific value can be determined according to the actual situation and is not limited. Based on this anomaly intensity coefficient, the weight vector q of the objective function is mapped to the Hessian matrix H using a piecewise linear function: when γ is below the first threshold γ1, the original weights remain unchanged; when γ is between γ1 and γ2, the penalty weights are increased linearly proportionally; when γ is above γ2, the weights are iteratively updated, triggering an exponential growth formula. Where β is the step size factor. In this way, the weights of the pump speed and fan speed control objectives are flexibly adjusted to form a new objective function. Where E(x) is the reconstruction error term and S(x) is the switching penalty term. and These are the penalty coefficients generated by mapping the anomaly intensity coefficient γ.

[0081] It should be noted that the switching penalty term in this application can constrain the abrupt changes in pump speed and fan speed, avoiding mechanical fatigue or electrical shock caused by frequent switching. For example, S(x) can be defined as... ,in This represents the difference between the control variables in the k-th period and the (k-1)-th period. By penalizing this term, control jitter under abnormal conditions can be effectively mitigated.

[0082] In the optimization process, the interior-point method is used to iteratively update the primal and dual variables. This process, centered on the KKT conditions, transforms the optimization problem into a system of coupled primal and dual equations. Let the primal variables be... The dual variable is The KKT system can then be represented as: Where A is the constraint matrix and b is the constraint vector, including duty cycle boundaries, rate limits, and integral limits. A Jacobian matrix J is constructed using a Newton direction search mechanism, and in each iteration, JΔz = −r is calculated to obtain the step size vector Δz, which is then used to update x and y. To prevent numerical divergence, an adaptive damping factor σ is introduced during the update process, making the update formula z(k+1) = z(k) + σΔz(k), where σ takes values ​​in (0,1) and is dynamically adjusted according to the residual descent rate.

[0083] In practical applications, if the pump speed duty cycle solution Reaching the boundary Then the projection operator is constrained by inequality. Limitation corrections are performed. Simultaneously, the dual variable is adjusted synchronously through the residual balancing function, for example... Where rp and rd are the original and dual residuals, respectively, and α is the equilibrium factor. This mechanism ensures that the dual variable maintains the KKT equilibrium condition after the original variable is adjusted, thereby improving the convergence stability of the interior point method.

[0084] In the weight iteration part, if γ > γc for w consecutive periods (γc is a preset threshold), then the exponential iteration update mode is entered, and the formula is: Where ρ is the growth scaling factor and λmax is the upper limit clamping value of the penalty coefficient. This method avoids the optimizer from losing solvability or incurring excessive penalties under long-term abnormal conditions, thereby keeping the system running within a controllable range.

[0085] By integrating the overall behavior of the charging pile liquid cooling control system, the penalty function mechanism of this collaborative scheduling module can automatically increase the weight of the corresponding error term when it detects coolant flow deviation, fan stall, or abnormal power fluctuation of the pump motor. This makes the optimization solution more inclined to reduce heat dissipation risks rather than pursue optimal energy efficiency. For example, when insufficient heat dissipation is caused by fan malfunction, the system increases the weight of the pump speed control term to prioritize coolant circulation capacity, thereby preventing overheating of the charging pile power devices. Conversely, when a decrease in pump efficiency is detected, the system enhances the fan speed control term through weight mapping to achieve air-side heat dissipation compensation.

[0086] It should be noted that the terms "anomaly intensity coefficient," "weight mapping," and "dual residual balance" in this application are all limiting expressions of the technical content of this application, used to illustrate how to tightly couple the detected dynamic anomaly signal with the mathematical optimization mechanism in a liquid cooling system, thereby achieving a more robust control strategy. Specific details can be determined according to the actual situation, and no limitation is imposed.

[0087] It should also be noted that the mathematical formulas, derivations, symbol definitions, and parameter calculation methods used in this specification are all for the purpose of further clarifying and demonstrating the technical content of this invention, so that those skilled in the art can more intuitively and accurately understand the working mechanism and technical effects of this invention. These formulas are only used as quantitative expressions or illustrative examples of technical features and do not constitute limiting conditions of the claims of this invention. Those skilled in the art should understand that, without changing the core idea of ​​this invention, the parameter forms, calculation methods, numerical ranges, and even symbol representations involved in the formulas can be equivalently replaced or simplified in engineering according to the actual application environment. The specifics can be determined according to the actual situation, and no limitation is imposed. It should also be emphasized that the formulas in this specification are not theoretical derivations in the style of academic research papers, but rather an engineering description of the embodiments of this invention. Their purpose is to enhance the understandability and implementability of this invention, rather than to increase redundancy and complexity. Those skilled in the art can choose whether to use such quantitative tools when reading this specification, or can achieve the same technical effects through other equivalent methods.

[0088] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A liquid-cooled heat dissipation control system for a charging pile; characterized in that: include: The data acquisition module is used to establish a unified time-series data stream based on pump current feedback, fan speed feedback and multi-point temperature sampling data through a time synchronization mechanism, and to generate a monitoring dataset using a weighted interpolation method, which is then output to the first prediction modeling module and the second baseline modeling module. The first predictive modeling module is used to generate a future heat load prediction curve based on the monitoring dataset using a time series modeling method, and to obtain the time delay compensation parameters of pumps and fans through cross-correlation phase processing. The prediction results and time delay parameters are output to the collaborative scheduling module in parallel. The second baseline modeling module is used to identify the relationship between circulating pump current, flow rate, fan speed and heat exchange efficiency online using the recursive least squares algorithm. During the identification process, an adaptive baseline model is constructed through a forgetting factor update mechanism, and the baseline prediction value and identification parameter set are output to the residual detection module and the collaborative scheduling module. The residual detection module is used to perform change point analysis based on the baseline predicted values ​​and the monitoring dataset, using a combined method of cumulative sum statistical test and generalized likelihood ratio detection, and output residual sequences, anomaly labels and confidence indices to the collaborative scheduling module. The collaborative scheduling module is used to construct a constrained quadratic programming optimization problem based on the outputs of the first prediction modeling module, the second baseline modeling module, and the residual detection module, and to obtain the joint control trajectory of the circulating pump and the fan by optimizing the solution using the interior point method. The joint control trajectory is output to the drive layer as the final execution command.

2. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The acquisition module is also used to: dynamically fuse the temperatures of different measuring points based on the spatial distribution weights of multi-point temperature sampling signals, using a combination of hierarchical weighted interpolation and time window sliding average, and construct a monitoring dataset with redundant verification fields under a unified time-series framework of pump current feedback and fan speed feedback.

3. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The time series modeling method includes: The charging power and coolant inlet and outlet temperature sequences of the monitoring data were detrended by sliding window mean to obtain a stationary time series. The stationary sequence is modeled based on an autoregressive moving average model, and the model parameters are solved by the maximum likelihood estimation method. In the prediction phase, the prediction results of the autoregressive moving average model are input into a Kalman smoother to correct for short-term noise fluctuations, resulting in a future heat load prediction curve; wherein, the future heat load prediction curve satisfy: （1） In formula (1), This is an autoregressive coefficient used to reflect the inheritance of historical heat load to future heat load; This is a sequence of historical heat load observations; This is the moving average coefficient, used to describe the corrective effect of historical residuals on predicted values; It is a historical residual sequence; For the residual term, For Kalman gain, The observation matrix; To monitor the dataset at the prediction step size The observation input; These are the prior predictions obtained based on the autoregressive moving average model; for .

4. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The processing method for cross-correlation phase processing includes: The pump current feedback sequence and fan speed feedback sequence in the monitoring data are normalized respectively, and the spectral features are obtained based on fast Fourier transform. The cross-power spectral density between pump current and fan speed is calculated in the frequency domain, and a sequence of cross-correlation functions is generated through normalization operations. Peak search is performed on the cross-correlation function sequence within a sliding time window to extract phase angle difference feature values, and the phase angle difference feature values ​​are mapped to time delay compensation parameters for the dynamic response of pumps and fans. The phase angle difference feature values ​​are used to update the cross-correlation function sequence based on time window sliding, so that the time delay compensation parameters are dynamically updated over time under conditions of photovoltaic power output fluctuations or charging power abrupt changes.

5. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The working process of the recursive least squares algorithm includes: Based on online sampling data of circulating pump current, flow rate, fan speed and heat exchange efficiency, an input vector and an output observation sequence are constructed. At the initial moment, the parameter estimation vector and covariance matrix are normalized to the identity matrix to generate the initial identification state; During the recursive process, the gain vector is calculated based on the input vector and the covariance matrix, the parameter estimation vector is updated successively, and the covariance matrix is ​​exponentially weighted and adjusted through the forgetting factor mechanism. During the parameter estimation vector update process, a threshold constraint limiting mechanism is used to adjust the parameter increment vector at each time step. The formula for component-level limiting is as follows: （2） In formula (2), and These represent the values ​​after the amplitude limiting process. The increment value of the first identification parameter and the first The identification parameters are in the first... The original increment value at time; Indicates that for the first The fixed threshold upper limit set by each parameter is used to reflect the sensitivity boundary of the pump current, flow rate or fan speed parameters to the heat exchange efficiency. The nonlinear adjustment index is a constant greater than 1, used to control the convergence rate of the amplitude limiting curve. For the adaptive limiting factor, in the first... Time based on residual variance Defined as: （3） In formula (3), This is the proportionality coefficient; when the residual fluctuation exceeds a preset residual variance threshold, Increase, the amplitude limit boundary expands; when the residual fluctuation is less than or equal to the residual variance threshold. Contraction, tightening of the amplitude limit boundary; Under the aforementioned component-level limiting processing, when hour, Keep parameter updates consistent with the original increment; when As the amplitude limiting function gradually converges, the parameter update results with controlled amplitude are obtained, generating a convergent and stable set of identification parameters.

6. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: During the identification process using the recursive least squares algorithm, the adaptive baseline model classifies the real-time correlation parameters of the circulating pump current, flow rate, fan speed, and heat exchange efficiency into three weighted groups corresponding to low-load, medium-load, and high-load operating conditions based on the ambient temperature, charging current, and coolant inlet temperature of the liquid-cooled circulation loop. In each sampling period, the weighted group is updated according to the joint evaluation index formed by the residual statistics and the forgetting factor. If the updated parameters continuously deviate beyond the preset threshold range under the same operating condition category, the parameter adjustment range of the weighted group of that category is constrained through a limiting constraint mechanism.

7. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The working process of the residual detection module is as follows: a multidimensional residual vector is constructed based on the residual sequence corresponding to the circulating pump current, coolant flow rate and fan speed, and the residual covariance matrix is ​​calculated in each detection window; When the eigenvalues ​​of the covariance matrix exceed the preset anomaly sensitivity range, a secondary discrimination procedure is triggered to compare the drift trend of the cumulative sum statistic with the likelihood gain function of the generalized likelihood ratio. If both satisfy the anomaly threshold condition, an anomaly label is generated and a confidence index is assigned. Meanwhile, when the residual detection module maintains the anomaly label state within the continuous detection window, it classifies the anomaly mode as a persistent operating condition offset and outputs the classification result to the collaborative scheduling module through the parameter feedback interface.

8. The liquid cooling heat dissipation control system for a charging pile according to claim 1, characterized in that: The collaborative scheduling module includes: Based on the heat load prediction curve output by the first prediction modeling module and the identification parameter set output by the second baseline modeling module, a quadratic programming optimization objective function is constructed. The optimization objective function is given duty cycle boundary constraints, rate limit constraints, and integral amplitude limit constraints to form a constrained convex optimization model. The constraints are transformed into a dual problem using the KKT conditions and the Lagrange multiplier method. The optimal solutions for pump speed and fan speed that satisfy the constraints are obtained by iteratively updating the original and dual variables using the interior point method. If the residual detection module outputs an abnormal label, the objective weights are dynamically corrected through the penalty function mechanism, and the quadratic programming objective function is solved again. The converged solution is transformed into a joint control trajectory of the circulating pump duty cycle and the fan speed, and then output to the drive layer.

9. The liquid cooling heat dissipation control system for a charging pile according to claim 8, characterized in that: The penalty function mechanism includes: Based on the anomaly labels and corresponding confidence indices output by the residual detection module, the anomaly intensity coefficient is calculated. The objective function weight vector and Hessian matrix are subjected to weight mapping processing. Weight parameters and penalty coefficients are generated from the anomaly intensity coefficients by piecewise linear mapping to obtain the reconstruction error term and switching penalty term. Based on the reconstructed error term, switching penalty term, and constraint set, the interior point method is used to resolve the quadratic programming problem to obtain intermediate control variables. If the anomaly intensity coefficient is greater than a preset threshold for w consecutive periods, then a weight iterative update is performed, the weight parameters are updated exponentially and an upper limit clamp is applied to the penalty coefficient to obtain a joint control trajectory.

10. A liquid cooling heat dissipation control system for a charging pile according to claim 8, characterized in that: The process of iteratively updating the original and dual variables using the interior point method includes: in each iteration step, constructing the original-dual coupled equations of pump speed and fan speed based on KKT conditions, solving the step size vector through the Newton direction search mechanism, and using an adaptive damping factor to perform numerical stability correction on the updated trajectory; during the update process, if the step size solution of the original variable touches the duty cycle boundary or the rate limit threshold, the variable is constrained and adjusted through the inequality constraint projection operator, while the dual variable is synchronously corrected through the residual balance function.

Images