An Adaptive Adjustment Method for Pulsating Cooling Demand Based on AI Load Forecasting

Abstract

This invention discloses a pulsed adaptive cooling demand adjustment method based on AI load forecasting, comprising: generating a load feature dataset in a unified format; obtaining a downsampled load feature dataset; sequentially extracting short-cycle fluctuation features, periodic trend features, and long-term evolution features; obtaining a calibration load uncertainty interval sequence; solving for the target cooling capacity and its optimal pulse duty cycle; converting the target cooling capacity into a pulsed start-stop signal to drive the compressor to start and stop within the pulse cycle according to the optimal pulse duty cycle; and immediately switching to a protection pulse mode to maintain safe system operation when the rate of change of the chilled water outlet temperature exceeds a safety threshold or a sensor self-check failure is detected. This invention retains detailed information about load abrupt changes, significantly improving the prediction accuracy and robustness of the model under non-stationary, multi-scale load environments.

Description

Technical Field

[0001] This invention relates to the field of fixed-frequency chiller technology, and in particular to a pulse-type adaptive adjustment method for cooling demand based on AI load prediction. Background Technology

[0002] In the energy management systems of modern large-scale public buildings and industrial facilities, central air conditioning systems account for a significant proportion of total building energy consumption. As the core of the cooling source, the operating efficiency of chillers directly determines the overall energy efficiency level of the system. Among them, fixed-frequency chillers, with their advantages of simple structure, low cost, and reliable operation, still occupy an important position in the existing market and certain new construction projects. However, fixed-frequency units can typically only operate at full speed at rated power or be completely shut down, unlike variable-frequency units which can continuously match the constantly changing cooling load demand at the terminal by adjusting the motor speed. This poses a significant challenge to optimizing their energy efficiency under partial load conditions.

[0003] In practical applications, existing load forecasting and control technologies for fixed-frequency chillers still have many shortcomings, mainly in the following aspects: Building cooling load data exhibits significant non-stationarity and multi-scale characteristics, including short-term random noise interference, obvious diurnal periodic fluctuations, and long-term evolution trends influenced by seasons. Existing load forecasting methods often employ single-scale forecasting models, failing to decouple and analyze different frequency components of the load sequence. This one-size-fits-all approach makes the model susceptible to being misled by high-frequency noise or unable to capture long-term, slow drifts, thus making it difficult to meet the requirements of refined control in terms of forecast accuracy.

[0004] Most existing scheduling strategies are based on deterministic point forecasts, meaning the model only outputs a specific load value. However, due to the randomness of factors such as weather changes, population movement, and equipment heat dissipation, a single forecast value cannot reflect future uncertainties. Once the actual load deviates significantly from the forecast value, scheduling schemes based on deterministic forecasts often fail, leading to insufficient cooling causing user discomfort, or excessive cooling resulting in energy waste. These strategies lack tolerance for forecast errors and risk quantification mechanisms.

[0005] When converting load demand into control commands for fixed-frequency units, existing pulse-type regulation methods often lack in-depth consideration of the mechanical characteristics of the compressor. As a precision mechanical device, the compressor has strict limitations on its minimum continuous operating time, minimum continuous shutdown time, and the number of start-stop cycles per unit time. Existing control logic often uses simple rule constraints and fails to jointly optimize and solve these physical constraints with dynamically changing load ranges. This can easily lead to conflicts between meeting load demand and protecting equipment lifespan, resulting in frequent start-stop cycles that cause equipment failure, or excessive protection that causes drastic water temperature fluctuations, making it impossible to achieve smooth and precise regulation of cooling capacity. Summary of the Invention

[0006] One objective of this invention is to propose a pulsed cooling demand adaptive adjustment method based on AI load forecasting. This invention retains detailed information about load abrupt changes and significantly improves the prediction accuracy and robustness of the model in non-stationary, multi-scale load environments.

[0007] According to an embodiment of the present invention, a pulse-type cooling capacity regulation method for a fixed-frequency chiller unit based on a hierarchical recursive prediction network and a load uncertainty interval constraint scheduling mechanism includes: Construct a multi-source operational data acquisition system, complete time synchronization processing, outlier removal processing, missing value imputation processing and normalization processing, and generate a load characteristic dataset in a unified format; Downsampling and noise suppression are performed on the load feature dataset at the edge to obtain a downsampled load feature dataset. Using a downsampled load feature dataset as input, a hierarchical recursive prediction network is constructed and trained offline. The hierarchical recursive prediction network encodes the load sequence hierarchically according to the inherent frequency decomposition strategy, and extracts short-cycle fluctuation features, periodic trend features and long-term evolution features in sequence. In the hierarchical recursive prediction network, the Monte Carlo random deactivation sampling method is used to output the mean sequence of future control time domain cold load prediction and the corresponding load uncertainty interval sequence. The load uncertainty interval sequence is then calibrated with confidence using a temperature calibration strategy to obtain the calibrated load uncertainty interval sequence. Based on the calibration load uncertainty interval sequence and the cooling load prediction mean sequence, a load uncertainty interval constraint rolling scheduling model is constructed. The calibration load uncertainty interval sequence is discretized into a scene tree through Latin hypercube sampling. Combining the minimum continuous running time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop times per unit time constraint, the target cooling capacity and its optimal pulse duty cycle are obtained. Based on the optimal pulse duty cycle, a compressor pulse start-stop control command set is generated within a preset pulse period, and the target cooling capacity is converted into a pulse start-stop signal to drive the compressor to start and stop within the pulse period according to the optimal pulse duty cycle. During the execution of the pulse-type start-stop control command set, the chilled water outlet temperature change rate data and compressor operating status data are monitored in real time. When the chilled water outlet temperature change rate data exceeds the safety threshold or the sensor self-test fails, the system immediately switches to the protection pulse mode to maintain safe operation.

[0008] Optionally, the construction of a multi-source operational data acquisition system, which completes time synchronization processing, outlier removal processing, missing value imputation processing, and normalization processing, generates a load characteristic dataset in a unified format, includes: Configure multiple sensor acquisition interfaces located at different physical nodes of the chiller unit system to read chilled water return temperature, chilled water outlet temperature, chilled water flow rate and ambient temperature and humidity data in parallel, forming a raw multi-source heterogeneous sensor data stream; A high-precision global clock reference based on a network time protocol is established. Using a preset minimum sampling time interval as a reference, timestamp alignment and interpolation resampling are performed on each data point in the original multi-source heterogeneous sensor data stream to eliminate the time deviation of data acquisition and generate a time-synchronized running data sequence. Using a statistical constant detection algorithm based on interquartile range, the upper and lower quartiles of each feature dimension in the time-synchronized running data sequence are calculated, outliers exceeding the statistical threshold are identified and removed, and a denoised running data sequence is generated. For missing positions in the denoised running data sequence caused by outlier removal or packet loss, a temporal interpolation model based on a long short-term memory network is adopted to predict and fill missing values ​​by utilizing the temporal dependencies before and after the missing position, thereby generating a complete running data sequence. Max-min normalization is performed on various physical quantity data in the integrity operation data sequence, and the normalized data is encapsulated in chronological order to generate a load characteristic dataset in a unified format.

[0009] Optionally, the downsampling and noise suppression processing of the load feature dataset at the edge side to obtain a downsampled load feature dataset includes: Input the load characteristic dataset in a unified format into the edge computing gateway, use the fast Fourier transform to analyze the spectral characteristics of the data, determine the redundancy multiple between the original sampling frequency and the system dynamic response frequency, and calculate the optimal downsampling factor. Based on the optimal downsampling factor, a multiphase polynomial anti-aliasing filter is designed to perform low-pass filtering on the load characteristic dataset, filter out high-frequency components above the Nyquist frequency, prevent spectral aliasing, and generate an anti-aliasing filtered data sequence. The data sequence of the anti-aliasing filter is extracted at equal intervals according to the optimal downsampling factor to generate a sparse load data sequence. A wavelet thresholding denoising algorithm is adopted. The Daubechies wavelet basis function is selected to perform multi-scale decomposition on the sparse load data sequence. The high-frequency detail coefficients after decomposition are shrunk by applying a soft thresholding function to remove the inherent white noise of the sensor. The smoothed data is generated by the reconstruction algorithm to obtain the downsampled load feature dataset.

[0010] Optionally, the hierarchical recursive prediction network performs hierarchical encoding on the load sequence according to an inherent frequency decomposition strategy, sequentially extracting short-cycle fluctuation features, periodic trend features, and long-term evolution features, including: Variational mode decomposition algorithm is introduced as the core of the intrinsic frequency decomposition strategy. The mode decomposition layer number parameter is set, and the downsampled load feature dataset is adaptively decomposed into several intrinsic mode function components with different center frequencies to form a multi-scale mode component set. Calculate the sample entropy of each component in the multi-scale modal component set, and automatically cluster each component into high-frequency random components, mid-frequency periodic components and low-frequency trend components according to the entropy value, generating a hierarchical frequency feature subset; A short-period extraction layer containing a one-dimensional convolutional neural network encoder is constructed to perform local perception and feature mapping on high-frequency random components in the hierarchical frequency feature subset, capture instantaneous load change information, and generate short-period fluctuation feature vectors. A periodic extraction layer containing gated recurrent units and attention mechanisms is constructed to perform time-series dependency modeling on the mid-frequency periodic components in the hierarchical frequency feature subset, and to weighted focus on key nodes in historical cycles to generate periodic trend feature vectors. A long-term evolution extraction layer containing a dilated convolutional network is constructed to extract low-frequency trend components from the hierarchical frequency feature subset with a large receptive field, capturing the overall drift and seasonal changes of the load, and generating long-term evolution feature vectors.

[0011] Optionally, the step of using Monte Carlo random deactivation sampling to output the mean sequence of future control time-domain cooling load forecasts and the corresponding load uncertainty interval sequence in the hierarchical recursive prediction network includes: The concatenated short-period fluctuation feature vector, periodic trend feature vector, and long-term evolution feature vector are input into the decoder layer of the hierarchical recursive prediction network, and the Dropout regularization module is enabled in the fully connected layer of the decoder. During the inference phase, the Dropout regularization module is kept active. A preset number of Monte Carlo sampling times N is set, and N forward propagation inferences are performed on the same input feature. In each inference, some neuron connections are randomly discarded to generate N sets of parallel cold load prediction sample sets. For each time step, the statistical mean of the N sets of cooling load prediction samples is calculated as a deterministic prediction output to generate a sequence of cooling load prediction mean values ​​for the future control time domain. Calculate the prediction variance of N sets of cooling load prediction samples at each time step. Based on the preset confidence level and Gaussian distribution assumption, use the formula of mean plus or minus k times the standard deviation to construct the preliminary upper and lower bound ranges and generate the original load uncertainty interval sequence.

[0012] Optionally, the step of performing confidence calibration on the load uncertainty interval sequence using a temperature calibration strategy to obtain a calibrated load uncertainty interval sequence includes: Using validation set data, the actual coverage rate of the original load uncertainty interval sequence at different confidence levels is calculated, and a reliability mapping curve between the predicted confidence level and the actual coverage rate is constructed. The area of ​​deviation between the reliability mapping curve and the ideal diagonal is calculated as a calibration error index. If the error index exceeds the threshold, a temperature scaling model based on isostatic regression is introduced. A scalar temperature parameter is learned using a temperature scaling model. This scalar temperature parameter is used to adjust the entropy value of the original prediction distribution, making the adjusted probability density of the distribution more consistent with the true error distribution, and generating a calibrated probability density function. Based on the calibrated probability density function, the quantile boundaries at the specified confidence level are recalculated, and the width and offset of the original interval are corrected to ensure that the interval can cover the true load value with nominal probability. Finally, a calibrated load uncertainty interval sequence with rigorous statistical calibration is obtained.

[0013] Optionally, the step of constructing a load uncertainty interval-constrained rolling scheduling model based on the calibration load uncertainty interval sequence and the cold load prediction mean sequence, and discretizing the calibration load uncertainty interval sequence into a scene tree using Latin hypercube sampling, includes: Extract the probability distribution characteristics of the calibration load uncertainty interval sequence at each time point in the future control time domain, and construct a multidimensional joint probability distribution model that varies with time. The Latin hypercube sampling algorithm is used to perform stratified sampling on the multidimensional joint probability distribution model to ensure that the sampling points uniformly cover the entire probability space and generate a large-scale set of initial load scenarios. The synchronous back-substitution reduction algorithm is used to calculate the probability distance between each scenario in the initial load scenario set, iteratively removes redundant scenarios and merges probabilities, retains the most representative scenarios, and generates the reduced load scenario set. The load scenario set is organized according to a time branching structure. The root node is defined as the current state, and the leaf nodes are the future load states under different probability paths. A multi-stage stochastic programming scenario tree that can characterize the uncertainty of future load evolution is constructed.

[0014] Optionally, the step of combining the minimum continuous operating time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop cycles per unit time constraint to obtain the target cooling capacity and its optimal pulse duty cycle includes: Based on a multi-stage stochastic programming scenario tree, a mixed-integer linear programming mathematical model is established with the objective of minimizing the weighted sum of operating energy consumption cost and chilled water temperature deviation penalty cost. In a mixed-integer linear programming mathematical model, binary state variables are introduced to represent the compressor's start-up and shutdown states, and a system of linear inequalities is constructed to describe the physical constraints. The minimum continuous operating time constraint forces the compressor to remain on for at least [duration missing]. The minimum continuous downtime constraint (minutes) forces the compressor to remain off once shut down for at least [number] minutes. minute; In the mixed-integer linear programming mathematical model, a sliding time window counter is introduced to limit the total number of start and stop actions within any sliding window to no more than a preset upper limit, thus forming a maximum start and stop constraint. The branch and bound algorithm is used to solve the mixed integer linear programming mathematical model with integrated constraints, and the optimal cooling power at the next moment under the weighted expectation of all scenarios is determined to generate the target cooling capacity command. Based on the ratio of the target cooling capacity command to the compressor's rated power, calculate the theoretical switching ratio that meets the current load demand and generate the optimal pulse duty cycle.

[0015] Optionally, the step of generating a compressor pulse-type start-stop control command set within a preset pulse period based on the optimal pulse duty cycle, and converting the target cooling capacity into a pulse-type start-stop signal, includes: Read the preset pulse width modulation reference period The reference period is divided into a high-level duration period and a low-level duration period; Multiply the optimal pulse duty cycle by the reference period. The specific number of seconds the compressor should remain running within a complete pulse cycle, i.e., the high-level duration, is calculated. And calculate the remaining downtime. ; Based on the duration of the high level With downtime Construct a digital logic sequence containing precise timestamps, which defines the setting of the output signal at the start of the cycle and... The timing logic for resetting the output signal at all times; The digital logic sequence is mapped to a 24V physical control voltage through the digital output port of the PLC controller, and hardware dead-time protection logic is superimposed to prevent signal jitter, generating a compressor pulse-type start-stop control instruction set for directly driving the compressor contactor to engage and disengage.

[0016] Optionally, during the execution of the pulse-type start-stop control command set, real-time monitoring of the chilled water outlet temperature change rate data and compressor operating status data includes: The instantaneous value of the chilled water outlet temperature is obtained in real time using a high-frequency temperature acquisition probe and stored in a first-in-first-out sliding data buffer. Apply first-order difference operation to the temperature data in the sliding data buffer and divide by the sampling time interval to calculate the real-time temperature change slope, i.e., the rate of change of chilled water outlet temperature. The absolute value of the chilled water outlet temperature change rate data is compared with the preset temperature fluctuation safety threshold. At the same time, the communication status between the temperature sensor and the compressor driver is polled through the heartbeat detection mechanism to generate a system safety status flag bit. When the system safety status flag indicates an abnormality, the current pulse-type start / stop control instruction set is interrupted, the protection logic is activated, and the preset fault safety parameter table is forcibly retrieved. Based on the fail-safe parameter table, an emergency control signal with a conservative duty cycle is generated and locked in this mode until manual reset or monitoring data returns to normal, thus generating the final system safe operation control flow.

[0017] The beneficial effects of this invention are: This invention employs a hybrid prediction architecture based on variational mode decomposition and hierarchical deep learning networks. By decoupling complex cold load sequences into intrinsic mode components of different frequencies, it achieves independent and accurate capture of short-period fluctuations, periodic trends, and long-term evolutionary features. The VMD algorithm is used to adaptively decompose the load data. For high-frequency components, a one-dimensional convolutional neural network is used to extract instantaneous mutation features. For mid-frequency components, a gated recurrent unit combined with an attention mechanism is used to capture periodic dependencies. For low-frequency components, a dilated convolutional network is used to extract long-term trends. This effectively eliminates the interference of noise on trend prediction while preserving the detailed information of load mutations, significantly improving the prediction accuracy and robustness of the model in non-stationary, multi-scale load environments.

[0018] This invention introduces an uncertainty quantification mechanism that combines Monte Carlo random deactivation sampling with a temperature calibration strategy. By constructing a statistically calibrated confidence interval based on the predicted mean, a reliable risk boundary is provided for scheduling decisions. By maintaining Dropout activation during the inference phase to generate multiple sets of predicted samples, the distribution variance of the load is calculated. Furthermore, a temperature calibration model based on equal-protection regression is used to calibrate the original uncertainty interval, correcting potential overconfidence or underconfidence issues in the model. This ensures that the output confidence interval can truly reflect the probability distribution of the prediction error, enabling the control system to perceive the future uncertainty range and thus reserve corresponding adjustment margins during scheduling. This effectively avoids frequent unit adjustments or temperature exceedances caused by prediction deviations.

[0019] This invention constructs a mixed-integer linear programming rolling scheduling model based on scenario trees. Combined with pulse width modulation (PWM) technology, it achieves quasi-continuous cooling capacity regulation of fixed-frequency units under strict constraints of compressor physical limitations. The calibrated load uncertainty interval is discretized into a multi-stage scenario tree. Physical constraints such as minimum continuous operating time, minimum continuous downtime, and maximum number of start-stop cycles of the compressor are explicitly embedded in the optimization model. By solving the MILP problem, the globally optimal cooling command is obtained and transformed into the optimal pulse duty cycle to drive the compressor. This not only mathematically guarantees the global optimality of the scheduling scheme and the safety of the equipment, but also transforms the discrete fixed-frequency output into a continuous cooling capacity output averaged over time using PWM. Without the need for inverter hardware investment, it significantly improves the energy efficiency ratio and temperature control accuracy of fixed-frequency units under partial load conditions, and extends the service life of the equipment. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a pulse-type adaptive adjustment method for cooling demand based on AI load forecasting proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 As shown in Example 1: A pulse-type cooling capacity regulation method for a fixed-frequency chiller unit based on a hierarchical recursive prediction network and a load uncertainty interval constraint scheduling mechanism, comprising: Construct a multi-source operational data acquisition system, complete time synchronization processing, outlier removal processing, missing value imputation processing and normalization processing, and generate a load characteristic dataset in a unified format; Downsampling and noise suppression are performed on the load feature dataset at the edge to obtain a downsampled load feature dataset. Using a downsampled load feature dataset as input, a hierarchical recursive prediction network is constructed and trained offline. The hierarchical recursive prediction network encodes the load sequence hierarchically according to the inherent frequency decomposition strategy, and extracts short-cycle fluctuation features, periodic trend features and long-term evolution features in sequence. In the hierarchical recursive prediction network, the Monte Carlo random deactivation sampling method is used to output the mean sequence of the future control time domain cooling load prediction and the corresponding load uncertainty interval sequence. The load uncertainty interval sequence is then calibrated with confidence using a temperature calibration strategy to obtain the calibrated load uncertainty interval sequence. Based on the calibration load uncertainty interval sequence and the cooling load prediction mean sequence, a load uncertainty interval constraint rolling scheduling model is constructed. The calibration load uncertainty interval sequence is discretized into a scene tree through Latin hypercube sampling. Combining the minimum continuous running time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop times per unit time constraint, the target cooling capacity and its optimal pulse duty cycle are obtained. Based on the optimal pulse duty cycle, a compressor pulse start-stop control command set is generated within a preset pulse cycle, and the target cooling capacity is converted into a pulse start-stop signal to drive the compressor to start and stop within the pulse cycle according to the optimal pulse duty cycle. During the execution of the pulse-type start-stop control command set, the chilled water outlet temperature change rate data and compressor operating status data are monitored in real time. When the chilled water outlet temperature change rate data exceeds the safety threshold or the sensor self-test failure is detected, the system immediately switches to the protection pulse mode to maintain safe operation.

[0023] In this embodiment, a multi-source operational data acquisition system is constructed to complete time synchronization processing, outlier removal processing, missing value imputation processing, and normalization processing, generating a load characteristic dataset in a unified format, including: Configure multiple sensor acquisition interfaces located at different physical nodes of the chiller unit system to read chilled water return temperature, chilled water outlet temperature, chilled water flow rate and ambient temperature and humidity data in parallel, forming a raw multi-source heterogeneous sensor data stream; A high-precision global clock reference based on a network time protocol is established. Using a preset minimum sampling time interval as a reference, timestamp alignment and interpolation resampling are performed on each data point in the original multi-source heterogeneous sensor data stream to eliminate the time deviation of data acquisition and generate a time-synchronized running data sequence. Using a statistical constant detection algorithm based on interquartile range, the upper and lower quartiles of each feature dimension in the time-synchronized running data sequence are calculated, outliers exceeding the statistical threshold are identified and removed, and a denoised running data sequence is generated. For missing positions in the denoised running data sequence caused by outlier removal or packet loss, a temporal interpolation model based on a long short-term memory network is adopted to predict and fill missing values ​​by utilizing the temporal dependencies before and after the missing position, thereby generating a complete running data sequence. Max-min normalization is performed on various physical quantity data in the integrity operation data sequence, and the normalized data is encapsulated in chronological order to generate a load characteristic dataset in a unified format.

[0024] In this embodiment, downsampling and noise suppression processing are performed on the load feature dataset at the edge side to obtain a downsampled load feature dataset, including: Input the load characteristic dataset in a unified format into the edge computing gateway, use the fast Fourier transform to analyze the spectral characteristics of the data, determine the redundancy multiple between the original sampling frequency and the system dynamic response frequency, and calculate the optimal downsampling factor. Based on the optimal downsampling factor, a multiphase polynomial anti-aliasing filter is designed to perform low-pass filtering on the load characteristic dataset, filter out high-frequency components above the Nyquist frequency, prevent spectral aliasing, and generate an anti-aliasing filtered data sequence. The data sequence of the anti-aliasing filter is extracted at equal intervals according to the optimal downsampling factor to generate a sparse load data sequence. A wavelet thresholding denoising algorithm is adopted. The Daubechies wavelet basis function is selected to perform multi-scale decomposition on the sparse load data sequence. The high-frequency detail coefficients after decomposition are shrunk by applying a soft thresholding function to remove the inherent white noise of the sensor. The smoothed data is generated by the reconstruction algorithm to obtain the downsampled load feature dataset.

[0025] In this embodiment, the hierarchical recursive prediction network encodes the load sequence hierarchically according to the inherent frequency decomposition strategy, sequentially extracting short-period fluctuation features, periodic trend features, and long-term evolution features, including: Variational mode decomposition algorithm is introduced as the core of the intrinsic frequency decomposition strategy. The mode decomposition layer number parameter is set, and the downsampled load feature dataset is adaptively decomposed into several intrinsic mode function components with different center frequencies to form a multi-scale mode component set. Calculate the sample entropy of each component in the multi-scale modal component set, and automatically cluster each component into high-frequency random components, mid-frequency periodic components and low-frequency trend components according to the entropy value, generating a hierarchical frequency feature subset; A short-period extraction layer containing a one-dimensional convolutional neural network encoder is constructed to perform local perception and feature mapping on high-frequency random components in the hierarchical frequency feature subset, capture instantaneous load change information, and generate short-period fluctuation feature vectors. A periodic extraction layer containing gated recurrent units and attention mechanisms is constructed to perform time-series dependency modeling on the mid-frequency periodic components in the hierarchical frequency feature subset, and to weighted focus on key nodes in historical cycles to generate periodic trend feature vectors. A long-term evolution extraction layer containing a dilated convolutional network is constructed to extract low-frequency trend components from the hierarchical frequency feature subset with a large receptive field, capturing the overall drift and seasonal changes of the load, and generating long-term evolution feature vectors.

[0026] In this embodiment, the Monte Carlo random deactivation sampling method is used in the hierarchical recursive prediction network to output the mean sequence of future control time-domain cooling load predictions and the corresponding load uncertainty interval sequence, including: The concatenated short-period fluctuation feature vector, periodic trend feature vector, and long-term evolution feature vector are input into the decoder layer of the hierarchical recursive prediction network, and the Dropout regularization module is enabled in the fully connected layer of the decoder. During the inference phase, the Dropout regularization module is kept active. A preset number of Monte Carlo sampling times N is set, and N forward propagation inferences are performed on the same input feature. In each inference, some neuron connections are randomly discarded to generate N sets of parallel cold load prediction sample sets. For each time step, the statistical mean of the N sets of cooling load prediction samples is calculated as a deterministic prediction output to generate a sequence of cooling load prediction mean values ​​for the future control time domain. Calculate the prediction variance of N sets of cooling load prediction samples at each time step. Based on the preset confidence level and Gaussian distribution assumption, use the formula of mean plus or minus k times the standard deviation to construct the preliminary upper and lower bound ranges and generate the original load uncertainty interval sequence.

[0027] In this embodiment, the load uncertainty interval sequence is calibrated with confidence using a temperature calibration strategy to obtain a calibrated load uncertainty interval sequence, including: Using validation set data, the actual coverage rate of the original load uncertainty interval sequence at different confidence levels is calculated, and a reliability mapping curve between the predicted confidence level and the actual coverage rate is constructed. The area of ​​deviation between the reliability mapping curve and the ideal diagonal is calculated as a calibration error index. If the error index exceeds the threshold, a temperature scaling model based on isostatic regression is introduced. A scalar temperature parameter is learned using a temperature scaling model. This scalar temperature parameter is used to adjust the entropy value of the original prediction distribution, making the adjusted probability density of the distribution more consistent with the true error distribution, and generating a calibrated probability density function. Based on the calibrated probability density function, the quantile boundaries at the specified confidence level are recalculated, and the width and offset of the original interval are corrected to ensure that the interval can cover the true load value with nominal probability. Finally, a calibrated load uncertainty interval sequence with rigorous statistical calibration is obtained.

[0028] In this embodiment, a load uncertainty interval constrained rolling scheduling model is constructed based on the calibration load uncertainty interval sequence and the cold load prediction mean sequence. The calibration load uncertainty interval sequence is discretized into a scene tree using Latin hypercube sampling, including: Extract the probability distribution characteristics of the calibration load uncertainty interval sequence at each time point in the future control time domain, and construct a multidimensional joint probability distribution model that varies with time. The Latin hypercube sampling algorithm is used to perform stratified sampling on the multidimensional joint probability distribution model to ensure that the sampling points uniformly cover the entire probability space and generate a large-scale set of initial load scenarios. The synchronous back-substitution reduction algorithm is used to calculate the probability distance between each scenario in the initial load scenario set, iteratively removes redundant scenarios and merges probabilities, retains the most representative scenarios, and generates the reduced load scenario set. The load scenario set is organized according to a time branching structure. The root node is defined as the current state, and the leaf nodes are the future load states under different probability paths. A multi-stage stochastic programming scenario tree that can characterize the uncertainty of future load evolution is constructed.

[0029] In this embodiment, by combining the minimum continuous operating time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop cycles per unit time constraint, the target cooling capacity and its optimal pulse duty cycle are obtained, including: Based on a multi-stage stochastic programming scenario tree, a mixed-integer linear programming mathematical model is established with the objective of minimizing the weighted sum of operating energy consumption cost and chilled water temperature deviation penalty cost. In a mixed-integer linear programming mathematical model, binary state variables are introduced to represent the compressor's start-up and shutdown states, and a system of linear inequalities is constructed to describe the physical constraints. The minimum continuous operating time constraint forces the compressor to remain on for at least [duration missing]. The minimum continuous downtime constraint (minutes) forces the compressor to remain off once shut down for at least [number] minutes. minute; In the mixed-integer linear programming mathematical model, a sliding time window counter is introduced to limit the total number of start and stop actions within any sliding window to no more than a preset upper limit, thus forming a maximum start and stop constraint. The branch and bound algorithm is used to solve the mixed integer linear programming mathematical model with integrated constraints, and the optimal cooling power at the next moment under the weighted expectation of all scenarios is determined to generate the target cooling capacity command. Based on the ratio of the target cooling capacity command to the compressor's rated power, calculate the theoretical switching ratio that meets the current load demand and generate the optimal pulse duty cycle.

[0030] In this embodiment, a compressor pulse-type start-stop control command set is generated within a preset pulse period based on the optimal pulse duty cycle, and the target cooling capacity is converted into a pulse-type start-stop signal, including: Read the preset pulse width modulation reference period The reference period is divided into a high-level duration period and a low-level duration period; Multiply the optimal pulse duty cycle by the reference period. The specific number of seconds the compressor should remain running within a complete pulse cycle, i.e., the high-level duration, is calculated. And calculate the remaining downtime. ; Based on the duration of the high level With downtime Construct a digital logic sequence containing precise timestamps, which defines the setting of the output signal at the start of the cycle and... The timing logic for resetting the output signal at all times; The digital logic sequence is mapped to a 24V physical control voltage through the digital output port of the PLC controller, and hardware dead-time protection logic is superimposed to prevent signal jitter, generating a compressor pulse-type start-stop control instruction set for directly driving the compressor contactor to engage and disengage.

[0031] In this embodiment, during the execution of the pulse-type start-stop control command set, the chilled water outlet temperature change rate data and compressor operating status data are monitored in real time, including: The instantaneous value of the chilled water outlet temperature is obtained in real time using a high-frequency temperature acquisition probe and stored in a first-in-first-out sliding data buffer. Apply first-order difference operation to the temperature data in the sliding data buffer and divide by the sampling time interval to calculate the real-time temperature change slope, i.e., the rate of change of chilled water outlet temperature. The absolute value of the chilled water outlet temperature change rate data is compared with the preset temperature fluctuation safety threshold. At the same time, the communication status between the temperature sensor and the compressor driver is polled through the heartbeat detection mechanism to generate a system safety status flag bit. When the system safety status flag indicates an abnormality, the current pulse-type start / stop control instruction set is interrupted, the protection logic is activated, and the preset fault safety parameter table is forcibly retrieved. Based on the fail-safe parameter table, an emergency control signal with a conservative duty cycle is generated and locked in this mode until manual reset or monitoring data returns to normal, thus generating the final system safe operation control flow.

[0032] Example 2: This example selects a centralized cooling system of a large public building as the application scenario. The building includes office areas, meeting areas, a data center, and a dining area. The total peak cooling load is approximately 3200kW, and it is equipped with three fixed-frequency chiller units, each with a rated cooling capacity of 1200kW. Since fixed-frequency chillers can only operate at full power or be completely shut down, there is a significant problem of over-cooling or frequent start-stops during off-peak conditions. Especially during the transitional periods between morning and evening and the peak midday hours with high pedestrian traffic, the cooling load exhibits strong pulse-like fluctuations, making it difficult for traditional control strategies to accurately match the load.

[0033] In the traditional control mode, the system employs a dual-threshold start-stop control logic based on the upper and lower limits of the chilled water outlet temperature. The compressor starts when the outlet temperature exceeds the set value of 7.5℃ and shuts down when it falls below 6.5℃. This control method is simple and direct, but it cannot predict load changes in advance. After analyzing actual operating data for 30 typical operating days, it was found that with a daily load fluctuation of 900kW, the compressor started and stopped an average of 42 times per day, with the shortest single operation time as low as 4 minutes, far below the recommended 10-minute continuous operation time. The standard deviation of the chilled water outlet temperature was 0.92℃, with a peak temperature fluctuation of 1.8℃ in some periods, and a user comfort complaint rate of 7.6%. The system's overall energy efficiency ratio (EER) was 4.05.

[0034] To address the aforementioned issues, the method of this invention is deployed and implemented in the building's cooling system. High-precision temperature sensors are installed at the return and outlet pipes of the chiller unit and the inlet of the cooling tower. Electromagnetic flow meters are installed on the main pipeline, and ambient temperature and humidity data are collected. The sampling period is set to 10 seconds. After 90 days of continuous data collection, a total of 777,600 raw data samples were obtained. After time synchronization, outlier removal, and missing data completion, the outlier rate was 2.7%, the packet loss rate was 1.3%, and the mean square error was controlled within 0.08 after LSTM interpolation completion. A unified load characteristic dataset is then formed after normalization.

[0035] Spectrum analysis was performed on the edge computing gateway. The original signal's main frequency was concentrated between 0.003Hz and 0.02Hz, while the original sampling frequency was 0.1Hz. The optimal downsampling factor was calculated to be 4, resulting in a sampling period of 40 seconds after downsampling. After multiphase anti-aliasing filtering and wavelet threshold denoising, the signal-to-noise ratio was improved from 28.6dB to 35.4dB.

[0036] A hierarchical recursive prediction network was constructed for offline training. 70% of the data was selected as the training set, 15% as the validation set, and 15% as the test set. Variational mode decomposition (VMD) was used to decompose the data into six modal components. After sample entropy clustering, three classes of components were formed: high-frequency, mid-frequency, and low-frequency. The short-cycle branch used a three-layer one-dimensional convolutional network, the mid-frequency branch used a two-layer GRU combined with an attention mechanism, and the low-frequency branch used a dilated convolutional network. The total number of parameters was approximately 1.85 × 10^6. The network converged after 120 training epochs.

[0037] The prediction results on the test set are as follows: the mean absolute error (MAE) of the traditional single-scale LSTM model is 86.4kW, the root mean square error (RMSE) is 124.7kW, and the MAPE is 6.9%; the MAE of the hierarchical recursive prediction network of this invention is 52.3kW, the RMSE is 73.5kW, and the MAPE is 4.1%. The prediction error decreased by 43% during high-frequency abrupt change periods.

[0038] During the uncertainty quantification phase, the number of Monte Carlo samplings was set to N=100. The initial 95% confidence interval actually covered 88.2%, indicating overconfidence. After temperature scaling calibration (temperature parameter T=1.37), the coverage improved to 94.6% after recalculation, and the calibration error area decreased from 0.084 to 0.019.

[0039] During midday on a certain operating day, the system load suddenly surged from 1600kW to 2350kW within 15 minutes. Traditional threshold control only activated the second unit when the temperature reached 7.6℃, resulting in a peak outlet water temperature of 8.3℃ and a recovery time of 22 minutes. The method of this invention predicts the load increase trend 10 minutes in advance from the load forecast mean sequence and provides a 95% confidence interval upper bound of 2420kW. The rolling scheduling model, after being discretized using a scenario tree, forms 50 representative scenarios. The mixed-integer linear programming solution time is 0.42 seconds, yielding a target cooling capacity of 2210kW and corresponding to an optimal pulse duty cycle of 0.84.

[0040] With the pulse cycle set to 300 seconds, then ton = 252 seconds and tof = 48 seconds. The compressor operates at this rhythm to achieve average cooling capacity matching over a time scale. Actual operating data shows that the highest outlet water temperature during this period was 7.4℃, the recovery time was 9 minutes, and the standard deviation of temperature fluctuation decreased to 0.38℃.

[0041] In a comparative test conducted continuously for 60 days, the statistical comparison between the method of this invention and the traditional control method is as follows: In terms of prediction accuracy, the MAE of this invention's model remains around 55kW, while the traditional model is around 90kW. Regarding energy consumption, the average daily power consumption decreased from 18650kWh to 17280kWh, achieving an energy saving rate of 7.35%. The number of compressor start-ups and shutdowns decreased from an average of 42 times per day to 18 times, and the shortest operating time increased to over 11 minutes. The standard deviation of chilled water outlet temperature decreased from 0.92℃ to 0.41℃. The equipment failure rate decreased from 2.3% to 0.8%. The user comfort complaint rate decreased to 1.2%.

[0042] For comparison of training samples, five consecutive days of load data were selected as examples. The traditional LSTM model had a 12-minute lag in predictions during the afternoon surge on the third day, with a maximum error of 210kW; the model of this invention had a lag of only 3 minutes, with a maximum error of 95kW. The uncertainty interval automatically widens by 18% during the surge period, providing a safety margin for scheduling.

[0043] In the safety protection test, the temperature sensor communication was interrupted. The system detected the abnormal heartbeat within 2 seconds and switched to the protection pulse mode. The duty cycle was locked at 0.6 to maintain safe operation and no temperature exceedance occurred.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A pulse-type adaptive adjustment method for cooling demand based on AI load forecasting, characterized in that, include: Construct a multi-source operational data acquisition system, complete time synchronization processing, outlier removal processing, missing value imputation processing and normalization processing, and generate a load characteristic dataset in a unified format; Downsampling and noise suppression are performed on the load feature dataset at the edge to obtain a downsampled load feature dataset. Using a downsampled load feature dataset as input, a hierarchical recursive prediction network is constructed and trained offline. The hierarchical recursive prediction network encodes the load sequence hierarchically according to the inherent frequency decomposition strategy, and extracts short-cycle fluctuation features, periodic trend features and long-term evolution features in sequence. In the hierarchical recursive prediction network, the Monte Carlo random deactivation sampling method is used to output the mean sequence of future control time domain cold load prediction and the corresponding load uncertainty interval sequence. The load uncertainty interval sequence is then calibrated with confidence using a temperature calibration strategy to obtain the calibrated load uncertainty interval sequence. Based on the calibration load uncertainty interval sequence and the cooling load prediction mean sequence, a load uncertainty interval constraint rolling scheduling model is constructed. The calibration load uncertainty interval sequence is discretized into a scene tree through Latin hypercube sampling. Combining the minimum continuous running time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop times per unit time constraint, the target cooling capacity and its optimal pulse duty cycle are obtained. Based on the optimal pulse duty cycle, a compressor pulse start-stop control command set is generated within a preset pulse period, and the target cooling capacity is converted into a pulse start-stop signal to drive the compressor to start and stop within the pulse period according to the optimal pulse duty cycle. During the execution of the pulse-type start-stop control command set, the chilled water outlet temperature change rate data and compressor operating status data are monitored in real time. When the chilled water outlet temperature change rate data exceeds the safety threshold or the sensor self-test fails, the system immediately switches to the protection pulse mode to maintain safe operation.

2. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, The aforementioned multi-source operational data acquisition system completes time synchronization processing, outlier removal processing, missing value imputation processing, and normalization processing to generate a unified format load characteristic dataset, including: Configure multiple sensor acquisition interfaces located at different physical nodes of the chiller unit system to read chilled water return temperature, chilled water outlet temperature, chilled water flow rate and ambient temperature and humidity data in parallel, forming a raw multi-source heterogeneous sensor data stream; A high-precision global clock reference based on a network time protocol is established. Using a preset minimum sampling time interval as a reference, timestamp alignment and interpolation resampling are performed on each data point in the original multi-source heterogeneous sensor data stream to eliminate the time deviation of data acquisition and generate a time-synchronized running data sequence. Using a statistical constant detection algorithm based on interquartile range, the upper and lower quartiles of each feature dimension in the time-synchronized running data sequence are calculated, outliers exceeding the statistical threshold are identified and removed, and a denoised running data sequence is generated. For missing positions in the denoised running data sequence caused by outlier removal or packet loss, a temporal interpolation model based on a long short-term memory network is adopted to predict and fill missing values ​​by utilizing the temporal dependencies before and after the missing position, thereby generating a complete running data sequence. Max-min normalization is performed on various physical quantity data in the integrity operation data sequence, and the normalized data is encapsulated in chronological order to generate a load characteristic dataset in a unified format.

3. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, The load feature dataset is downsampled and noise suppressed at the edge to obtain a downsampled load feature dataset, including: Input the load characteristic dataset in a unified format into the edge computing gateway, use the fast Fourier transform to analyze the spectral characteristics of the data, determine the redundancy multiple between the original sampling frequency and the system dynamic response frequency, and calculate the optimal downsampling factor. Based on the optimal downsampling factor, a multiphase polynomial anti-aliasing filter is designed to perform low-pass filtering on the load characteristic dataset, filter out high-frequency components above the Nyquist frequency, prevent spectral aliasing, and generate an anti-aliasing filtered data sequence. The data sequence of the anti-aliasing filter is extracted at equal intervals according to the optimal downsampling factor to generate a sparse load data sequence. A wavelet thresholding denoising algorithm is adopted. The Daubechies wavelet basis function is selected to perform multi-scale decomposition on the sparse load data sequence. The high-frequency detail coefficients after decomposition are shrunk by applying a soft thresholding function to remove the inherent white noise of the sensor. The smoothed data is generated by the reconstruction algorithm to obtain the downsampled load feature dataset.

4. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, The hierarchical recursive prediction network encodes the load sequence hierarchically according to an inherent frequency decomposition strategy, sequentially extracting short-period fluctuation features, periodic trend features, and long-term evolution features, including: Variational mode decomposition algorithm is introduced as the core of the intrinsic frequency decomposition strategy. The mode decomposition layer number parameter is set, and the downsampled load feature dataset is adaptively decomposed into several intrinsic mode function components with different center frequencies to form a multi-scale mode component set. Calculate the sample entropy of each component in the multi-scale modal component set, and automatically cluster each component into high-frequency random components, mid-frequency periodic components and low-frequency trend components according to the entropy value, generating a hierarchical frequency feature subset; A short-period extraction layer containing a one-dimensional convolutional neural network encoder is constructed to perform local perception and feature mapping on high-frequency random components in the hierarchical frequency feature subset, capture instantaneous load change information, and generate short-period fluctuation feature vectors. A periodic extraction layer containing gated recurrent units and attention mechanisms is constructed to perform time-series dependency modeling on the mid-frequency periodic components in the hierarchical frequency feature subset, and to weighted focus on key nodes in historical cycles to generate periodic trend feature vectors. A long-term evolution extraction layer containing a dilated convolutional network is constructed to extract low-frequency trend components from the hierarchical frequency feature subset with a large receptive field, capturing the overall drift and seasonal changes of the load, and generating long-term evolution feature vectors.

5. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, The method of using Monte Carlo random deactivation sampling in the hierarchical recursive prediction network to output the mean sequence of future control time-domain cooling load predictions and the corresponding load uncertainty interval sequence includes: The concatenated short-period fluctuation feature vector, periodic trend feature vector, and long-term evolution feature vector are input into the decoder layer of the hierarchical recursive prediction network, and the Dropout regularization module is enabled in the fully connected layer of the decoder. During the inference phase, the Dropout regularization module is kept active. A preset number of Monte Carlo sampling times N is set, and N forward propagation inferences are performed on the same input feature. In each inference, some neuron connections are randomly discarded to generate N sets of parallel cold load prediction sample sets. For each time step, the statistical mean of the N sets of cooling load prediction samples is calculated as a deterministic prediction output to generate a sequence of cooling load prediction mean values ​​for the future control time domain. Calculate the prediction variance of N sets of cooling load prediction samples at each time step. Based on the preset confidence level and Gaussian distribution assumption, use the formula of mean plus or minus k times the standard deviation to construct the preliminary upper and lower bound ranges and generate the original load uncertainty interval sequence.

6. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 5, characterized in that, The step of calibrating the load uncertainty interval sequence using a temperature calibration strategy to obtain a calibrated load uncertainty interval sequence includes: Using validation set data, the actual coverage rate of the original load uncertainty interval sequence at different confidence levels is calculated, and a reliability mapping curve between the predicted confidence level and the actual coverage rate is constructed. The area of ​​deviation between the reliability mapping curve and the ideal diagonal is calculated as a calibration error index. If the error index exceeds the threshold, a temperature scaling model based on isostatic regression is introduced. A scalar temperature parameter is learned using a temperature scaling model. This scalar temperature parameter is used to adjust the entropy value of the original prediction distribution, making the adjusted probability density of the distribution more consistent with the true error distribution, and generating a calibrated probability density function. Based on the calibrated probability density function, the quantile boundaries at the specified confidence level are recalculated, and the width and offset of the original interval are corrected to ensure that the interval can cover the true load value with nominal probability. Finally, a calibrated load uncertainty interval sequence with rigorous statistical calibration is obtained.

7. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 6, characterized in that, The load uncertainty interval constraint rolling scheduling model is constructed based on the calibration load uncertainty interval sequence and the cold load prediction mean sequence. The calibration load uncertainty interval sequence is discretized into a scene tree using Latin hypercube sampling, including: Extract the probability distribution characteristics of the calibration load uncertainty interval sequence at each time point in the future control time domain, and construct a multidimensional joint probability distribution model that varies with time. The Latin hypercube sampling algorithm is used to perform stratified sampling on the multidimensional joint probability distribution model to ensure that the sampling points uniformly cover the entire probability space and generate a large-scale set of initial load scenarios. The synchronous back-substitution reduction algorithm is used to calculate the probability distance between each scenario in the initial load scenario set, iteratively removes redundant scenarios and merges probabilities, retains the most representative scenarios, and generates the reduced load scenario set. The load scenario set is organized according to a time branching structure. The root node is defined as the current state, and the leaf nodes are the future load states under different probability paths. A multi-stage stochastic programming scenario tree that can characterize the uncertainty of future load evolution is constructed.

8. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 7, characterized in that, The method combines the minimum continuous operating time constraint of the compressor, the minimum continuous shutdown time constraint of the compressor, and the maximum number of start-stop cycles per unit time constraint to solve for the target cooling capacity and its optimal pulse duty cycle, including: Based on a multi-stage stochastic programming scenario tree, a mixed-integer linear programming mathematical model is established with the objective of minimizing the weighted sum of operating energy consumption cost and chilled water temperature deviation penalty cost. In a mixed-integer linear programming mathematical model, binary state variables are introduced to represent the compressor's start-up and shutdown states, and a system of linear inequalities is constructed to describe the physical constraints. The minimum continuous operating time constraint forces the compressor to remain on for at least [duration missing]. The minimum continuous downtime constraint (minutes) forces the compressor to remain off once shut down for at least [number] minutes. minute; In the mixed-integer linear programming mathematical model, a sliding time window counter is introduced to limit the total number of start and stop actions within any sliding window to no more than a preset upper limit, thus forming a maximum start and stop constraint. The branch and bound algorithm is used to solve the mixed integer linear programming mathematical model with integrated constraints, and the optimal cooling power at the next moment under the weighted expectation of all scenarios is determined to generate the target cooling capacity command. Based on the ratio of the target cooling capacity command to the compressor's rated power, calculate the theoretical switching ratio that meets the current load demand and generate the optimal pulse duty cycle.

9. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, The step of generating a compressor pulse-type start-stop control command set within a preset pulse period based on the optimal pulse duty cycle, and converting the target cooling capacity into a pulse-type start-stop signal, includes: Read the preset pulse width modulation reference period The reference period is divided into a high-level duration period and a low-level duration period; Multiply the optimal pulse duty cycle by the reference period. The specific number of seconds the compressor should remain running within a complete pulse cycle, i.e., the high-level duration, is calculated. And calculate the remaining downtime. ; Based on the duration of the high level With downtime Construct a digital logic sequence containing precise timestamps, which defines the setting of the output signal at the start of the cycle and... The timing logic for resetting the output signal at all times; The digital logic sequence is mapped to a 24V physical control voltage through the digital output port of the PLC controller, and hardware dead-time protection logic is superimposed to prevent signal jitter, generating a compressor pulse-type start-stop control instruction set for directly driving the compressor contactor to engage and disengage.

10. The pulse-type adaptive adjustment method for cooling demand based on AI load forecasting according to claim 1, characterized in that, During the execution of the pulse-type start-stop control command set, the real-time monitoring of chilled water outlet temperature change rate data and compressor operating status data includes: The instantaneous value of the chilled water outlet temperature is obtained in real time using a high-frequency temperature acquisition probe and stored in a first-in-first-out sliding data buffer. Apply first-order difference operation to the temperature data in the sliding data buffer and divide by the sampling time interval to calculate the real-time temperature change slope, i.e., the rate of change of chilled water outlet temperature. The absolute value of the chilled water outlet temperature change rate data is compared with the preset temperature fluctuation safety threshold. At the same time, the communication status between the temperature sensor and the compressor driver is polled through the heartbeat detection mechanism to generate a system safety status flag bit. When the system safety status flag indicates an abnormality, the current pulse-type start / stop control instruction set is interrupted, the protection logic is activated, and the preset fault safety parameter table is forcibly retrieved. Based on the fail-safe parameter table, an emergency control signal with a conservative duty cycle is generated and locked in this mode until manual reset or monitoring data returns to normal, thus generating the final system safe operation control flow.

Images