A method and system for controlling air conditioning chilled water based on deep learning

Abstract

The present invention relates to the field of air-conditioning chilled water control technology, specifically to an air-conditioning chilled water control method and system based on deep learning, comprising the following steps: collecting the status of the air-conditioning chiller, including water pump frequency and valve opening feedback, as well as the pipe network supply water temperature, return water temperature, flow rate and pressure difference operation readings. The present invention dynamically collects the chiller status parameters and constructs a system operation status feature set, thereby realizing the real-time fusion and standardized processing of multi-source heterogeneous data, eliminating the timing misalignment problem caused by differences in sensor sampling frequencies, identifying the key input quantities of heat load changes based on the feature set, accurately predicting the performance attenuation trend of the water pump in combination with predictive calculations, and reducing the additional energy consumption caused by hysteresis regulation. A hierarchical optimization framework with the goal of minimizing total energy consumption is introduced, and the water supply temperature, pressure difference setting and unit operation combination are decoupled into independent optimization sub-problems, thereby reducing the ineffective work of the water pump while meeting the terminal heat demand.

Description

Technical Field

[0001] The present invention relates to the technical field of air-conditioning chilled water control, and in particular to an air-conditioning chilled water control method and system based on deep learning. Background Art

[0002] The field of air conditioning chilled water control technology focuses on achieving precise control of building environment temperature and humidity by adjusting the operating parameters of the chilled water system (temperature, pressure, flow, etc.). Its core covers the coordinated regulation of chillers, water pumps, valves, heat exchangers and terminal equipment.

[0003] Existing static control strategies frequently start and stop equipment under partial load conditions, increasing pump energy consumption and exacerbating mechanical wear. They also fail to account for the pump's efficiency decay curve. Chillers must be shut down and reset when output deviation exceeds a threshold, resulting in reduced stability. Safety constraints employ fixed thresholds to limit adjustment, making it difficult to dynamically adapt to real-time operating conditions. This increases the probability of triggering protective shutdowns during extreme load fluctuations. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and propose an air-conditioning chilled water control method and system based on deep learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an air conditioning chilled water control method based on deep learning, comprising the following steps:

[0006] Collect the status of the air-conditioning chiller, including pump frequency and valve opening feedback, as well as the pipe network supply water temperature, return water temperature, flow rate and pressure differential operation readings, synchronize and structure the signals, and establish a system operation status feature set;

[0007] Based on the system operation state feature set, key input quantities that affect heat load changes are identified to obtain prediction calculation input features, and based on the prediction calculation input features, prediction calculations are performed to obtain the performance status of the chiller water pump to obtain a system state prediction value;

[0008] Combining the system state prediction value with the current system operating state feature set, a calculation framework with the goal of minimizing total energy consumption is constructed and associated operating boundaries are set to obtain a description of the system energy efficiency optimization problem. Based on the description of the system energy efficiency optimization problem, a solution is performed to determine the water supply temperature, pressure difference set value and chiller operation combination, and establish hierarchical optimization control setting parameters;

[0009] By comparing the target setting values ​​in the hierarchical optimization control setting parameters with the real-time operating values ​​fed back by the system operating status feature set, the required adjustment amount of the water pump frequency is calculated to obtain the equipment adjustment difference calculation result. Based on the equipment adjustment difference calculation result, control instructions are generated according to the equipment response characteristics and safety constraints and organized in time sequence to obtain a real-time equipment adjustment instruction sequence.

[0010] Preferably, the steps of obtaining the system operation status feature set are:

[0011] Collect the pump frequency value, valve opening feedback value, pipe network supply water temperature value, return water temperature value, flow value and pressure difference operation reading value of the air conditioning chiller. All values ​​under the same device code are aligned to the system global clock reference at the time of collection. After marking the device code and timestamp, the original operating parameter set is generated;

[0012] Based on the original operating parameter set, for values ​​with missing timestamps or clock offsets between devices, the cubic spline interpolation method is used to complete the data sequence at 0.5-second time intervals. The temperature and flow values ​​are linearly normalized respectively. The sliding average filtering method with a window width of 5 data points is used to eliminate instantaneous fluctuation noise, and a system operating status feature set containing device codes, correction timestamps and denoising parameters is generated.

[0013] Preferably, the step of obtaining the prediction calculation input features is:

[0014] Based on the system operating status feature set, extract the time series data of the supply water temperature, return water temperature, flow rate, differential pressure operation reading, pump frequency, and valve opening feedback value. Statistically analyze the frequency of occurrence of the temperature value at 0.5°C intervals and the fluctuation amplitude of the flow rate value at 1m³ / h intervals. Calculate the Pearson linear correlation coefficient between each bin parameter and the historical heat load value to generate a parameter-heat load correlation coefficient set.

[0015] Based on the parameter-heat load correlation coefficient set, parameters with an absolute value of the Pearson linear correlation coefficient greater than 0.5 are screened, the time series lag relationship with the heat load value is verified, and a candidate set of key input quantities is generated;

[0016] Based on the candidate set of key input quantities, principal component analysis is performed on the supply water temperature values ​​and the return water temperature values ​​to eliminate multicollinearity, the variance inflation factors of the remaining parameters are calculated, and the parameters with variance inflation factors greater than 5 are eliminated. A calculation factor matrix including the orthogonalized temperature principal component, flow value and pressure difference operation reading value is constructed to generate the prediction calculation input features.

[0017] Preferably, the steps for obtaining the system state prediction value are:

[0018] Based on the predicted calculation input features, the orthogonalized temperature principal component values ​​are extracted , flow value and differential pressure operating readings , perform time series processing on each parameter with a 2-hour lag and fill the missing values ​​with the moving average of the first 3 data points to generate a time series parameter matrix ;

[0019] Based on the time series parameter matrix , establish an ARIMA model, train the model parameters by minimizing the mean square error between the predicted value and the actual heat load value, and generate a trained prediction model;

[0020] Based on the trained prediction model, the system state prediction value within the next 3 hours is calculated.

[0021] Preferably, the steps for obtaining the description of the system energy efficiency optimization problem are:

[0022] Combine the system state prediction value with the supply water temperature value, return water temperature value, flow value and pressure difference operation reading value in the system operation state feature set, align the predicted data for the next 3 hours with the actual data at the current moment by timestamp, verify the data integrity and fill in the missing values ​​with the average of two adjacent time points to generate a joint state time series data set;

[0023] Based on the joint state time series data set, the total energy consumption calculation objective function is defined as a linear combination of the cube of the water pump frequency value and the square of the flow value. Constraints are set, including that the difference between the supply water temperature and the return water temperature does not exceed 5°C and the fluctuation range of the pressure difference operating reading does not exceed ±10% of the current value. The objective function and constraint condition set are generated;

[0024] Based on the objective function and the set of constraints, a linear programming model is constructed, and the optimization variables are determined to be the water pump frequency adjustment amount, the valve opening adjustment amount and the start and stop status of the chiller, which are integrated into a system energy efficiency optimization problem description including the objective function, constraints and optimization variables.

[0025] Preferably, the step of obtaining the hierarchical optimization control setting parameters is:

[0026] Based on the objective function and constraint set in the system energy efficiency optimization problem description, the optimization range of the supply water temperature set point is limited to 15.0°C to 20.0°C. A list of temperature candidate values ​​is generated at 0.5°C intervals. For each candidate value, the corresponding water pump frequency adjustment amount and the return water temperature difference are calculated. Candidate values ​​that exceed the return water temperature difference limit of 5°C are eliminated to generate a set of feasible candidate values ​​for the supply water temperature.

[0027] Based on the set of feasible candidate values ​​for the water supply temperature, traverse the pressure differential set value range of 200 kPa to 500 kPa, verify whether the flow value fluctuation corresponding to each candidate pressure differential value is within the allowable range of ±10%, record the candidate pressure differential values ​​that meet the flow constraint and the corresponding chiller start and stop state combination codes, and generate a joint feasible solution set of pressure differential-unit combinations;

[0028] Based on the set of feasible solutions for the pressure difference-unit combination, the top 10% of the optimal solutions are selected in ascending order of the objective function value, the minimum start-stop interval time and the maximum daily start-stop limit of the chiller start-stop state combination are verified, the water supply temperature set value, the pressure difference set value and the chiller start-stop code sequence that meets the operation safety constraints are determined, and the hierarchical optimization control setting parameters are generated.

[0029] Preferably, the steps for obtaining the equipment adjustment difference calculation result are:

[0030] Extract the water supply temperature setting value, pressure difference setting value and chiller start / stop code from the hierarchical optimization control setting parameters, synchronously obtain the corresponding real-time water supply temperature, real-time pressure difference and water pump frequency in the system operation status feature set, calculate the temperature deviation accumulation and pressure difference change rate at three consecutive time points according to the timestamp, and generate a dynamic deviation parameter set;

[0031] Based on the dynamic deviation parameter set, the allowable boundary threshold of the water pump frequency adjustment amount is calculated, and the calculation formula is:

[0032] ;

[0033] in, is the boundary threshold, is the reference proportional coefficient of the pump frequency and temperature deviation, For the The temperature change rate at a time point, The difference between the current pressure difference set value and the real-time value, is the pressure difference reference constant, Number the current time point;

[0034] Based on the boundary threshold, the difference between the target frequency adjustment amount and the real-time frequency is limited to Within the range, the adjustment amount that exceeds the threshold is truncated according to the boundary value, and the device adjustment difference calculation result is generated, which includes the timestamp, the truncated adjustment amount and the constraint logic.

[0035] Preferably, the steps of acquiring the real-time device adjustment instruction sequence are:

[0036] Extracting a timestamp adjustment amount sequence from the equipment adjustment difference calculation result, calculating the absolute value mean and standard deviation of each adjustment amount according to the time window, eliminating abnormal adjustments that exceed a range of 3 times the standard deviation, and generating a cleaned adjustment amount sequence;

[0037] Based on the adjustment amount sequence after cleaning, the delay effect factor of the control instruction is calculated, and the calculation formula is:

[0038] ;

[0039] in, is the delay factor, is the end index of the current time window, For the The amount of adjustment after cleaning at each time point, is the device response delay base, is the instruction decay time constant;

[0040] Based on the delay effectiveness factor, delay compensation is added to each adjustment amount, and it is checked whether the compensated instruction exceeds the maximum step change rate of the water pump frequency, and a real-time equipment adjustment instruction sequence is generated.

[0041] The present invention provides an air-conditioning chilled water control system, comprising:

[0042] The status data acquisition module collects the status of the chiller, obtains water pump frequency feedback, valve opening feedback, pipe network supply water temperature reading, return water temperature reading, flow reading and pressure difference operation reading, performs signal synchronization processing, performs signal structure processing, and establishes synchronized structured operation reading;

[0043] The operation feature extraction module establishes a system operation status feature set based on the synchronous structured operation readings, identifies the key input quantities that affect the heat load change, obtains the input features required for the prediction calculation, and establishes the system operation status feature set;

[0044] The unit performance prediction module selects prediction calculation input features based on the system operating status feature set, performs prediction operations, obtains the performance status of the chiller pump, evaluates the expected interaction between the supply water temperature setting and the cooling tower fan speed adjustment under different loads, determines the expected system pressure drop of the evaporator and condenser, and obtains the system status prediction value;

[0045] The energy efficiency optimization setting module combines the system state prediction value with the current system operation state feature set to build a calculation framework with the goal of minimizing total energy consumption. It sets the associated operation boundaries, considers the capacity limit of the chiller and the pressure requirement of the pipe network, solves the system energy efficiency optimization problem description, evaluates different water supply temperature set points, pressure difference set points and chiller operation combination schemes, determines the optimal water supply temperature set point, pressure difference set point and chiller operation combination, and establishes the hierarchical optimization control setting parameters.

[0046] The equipment instruction generation module compares the target setting values ​​in the hierarchical optimization control setting parameters with the real-time operating values ​​fed back by the system operation status feature set, calculates the required adjustment amount of the water pump frequency, and obtains the equipment adjustment difference calculation result. Based on the equipment adjustment difference calculation result, the control instruction is generated according to the equipment response characteristics to obtain a real-time equipment adjustment instruction sequence.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are:

[0048] This invention dynamically collects chiller status parameters and constructs a system operating status feature set, enabling real-time fusion and standardized processing of heterogeneous data from multiple sources. This eliminates timing misalignment caused by differences in sensor sampling frequencies. Key input variables associated with heat load changes are identified based on the feature set, and combined with predictive calculations, this approach accurately predicts pump performance degradation trends, reducing the additional energy consumption caused by delayed regulation. This system introduces a hierarchical optimization framework with the goal of minimizing total energy consumption, decoupling the water supply temperature, pressure differential settings, and unit operation into independent optimization subproblems. This approach reduces ineffective pump work while still meeting terminal heat demand. Control commands are dynamically generated based on the calculated device adjustment differential. Delay compensation and safety threshold constraints are implemented based on device response characteristics to avoid overshoot and shorten adjustment time. Fluctuations in pipeline network pressure differentials are controlled within ±5 kPa of the set value. Real-time device adjustment command sequences are prioritized for execution, resolving resource contention conflicts when multiple devices operate in concert, minimizing synchronization errors between valve opening and pump frequency adjustments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] See also Figure 1 The present invention provides a technical solution, a deep learning-based air conditioning chilled water control method, comprising the following steps:

[0052] Collect the status of the air-conditioning chiller, including pump frequency and valve opening feedback, as well as the pipe network supply water temperature, return water temperature, flow rate and pressure differential operation readings, synchronize and structure the signals, and establish a system operation status feature set;

[0053] Based on the system operation status feature set, the key input quantities that affect the heat load change are identified to obtain the prediction calculation input features. Based on the prediction calculation input features, the prediction calculation is performed to obtain the performance status of the chiller water pump and obtain the system status prediction value;

[0054] Combining the system state prediction value with the current system operation state feature set, a calculation framework with the goal of minimizing total energy consumption is constructed and associated operation boundaries are set. This leads to a description of the system energy efficiency optimization problem. Based on this description of the system energy efficiency optimization problem, a solution is performed to determine the water supply temperature, pressure difference set values, and chiller operation combinations, and to establish hierarchical optimization control setting parameters.

[0055] By comparing the target set values ​​in the hierarchical optimization control setting parameters with the real-time operating values ​​fed back by the system operating status feature set, the required adjustment amount of the water pump frequency is calculated to obtain the equipment adjustment difference calculation result. Based on the equipment adjustment difference calculation result, control instructions are generated according to the equipment response characteristics and safety constraints and organized in time sequence to obtain a real-time equipment adjustment instruction sequence.

[0056] The steps for obtaining the system operation status feature set are:

[0057] Collect the pump frequency value, valve opening feedback value, pipe network supply water temperature value, return water temperature value, flow value and pressure difference operation reading value of the air conditioning chiller. All values ​​under the same device code are aligned to the system global clock reference at the time of collection. After marking the device code and timestamp, the original operating parameter set is generated;

[0058] Based on the original operating parameter set, for values ​​with missing timestamps or clock offsets between devices, the cubic spline interpolation method is used to complete the data sequence at 0.5-second time intervals. The temperature and flow values ​​are linearly normalized respectively, and the sliding average filtering method with a window width of 5 data points is used to eliminate instantaneous fluctuation noise. The system operating status feature set containing device codes, correction timestamps and denoising parameters is generated.

[0059] Specifically, data from various sensors deployed in the air-conditioning chiller are collected, including frequency readings output by the water pump inverter, feedback signals from electric valve strokes installed on each terminal or bypass pipeline, readings of supply water temperature sensors installed on the chilled water main, readings of return water temperature sensors installed on the chilled water main, readings of flow meters installed on the main pipeline or parallel branch, and readings of differential pressure sensors arranged between the supply and return water mains. These values ​​from different physical locations and measurement units are aggregated to the central data processing server through industrial fieldbus protocols, such as ModbusTCP or BACnet / IP protocols. The server assigns a unique device code to each independent sensor or device. For example, the supply water temperature sensor is coded as T_Supply_Main, the frequency of pump No. 1 is coded as Pump_Freq_01, and the opening of the two-way valve in area A is coded as Valve_Open_A2. When receiving the data packet, the original measurement value and the timestamp information of the device are first parsed. Then, according to the built-in Network Time Protocol (NTP) client of the central data processing server and the preset The system periodically synchronizes (for example, every minute) with the precise time obtained from the enterprise's internal NTP server (for example, with the IP address 192.168.1.10). The timestamp of each data record is forcibly corrected to the time based on this global unified clock reference. Specifically, the system reads the receiving timestamp (server time) of the data packet and uses it to overwrite or correct the device's native timestamp. The maximum allowed deviation is set to 500 milliseconds. If the device timestamp deviates from the server timestamp by more than 500 milliseconds, a warning log is logged, but the server timestamp still prevails. All value records with the same device code are then sorted according to the corrected global timestamp and organized into a time series format. For each data point, the corresponding device code, such as Pump_Freq_01, and the corrected precise timestamp are clearly marked in the format of YYYY-MM-DDHH:MM:SS.sss. Finally, the values ​​of all devices at all collection times, with the device code and corrected timestamp, are combined to form a structured data set, which is the original operating parameter set.

[0060] Based on the original operating parameter set obtained in the previous step, first perform a time series integrity check and data filling, traverse the time series data under each device code, and find out whether there is a situation where the timestamp interval is greater than the preset basic sampling interval (for example, set to 0.5 seconds), or whether there is a significant offset in the timestamps of different devices at the same theoretical sampling time (for example, the offset is greater than 100 milliseconds). Identify the data points with missing timestamps or clock offsets. For the identified missing points or points that need to be aligned, call the cubic spline interpolation algorithm to reconstruct and fill the data series. The algorithm uses two known data points before and after the missing point (a total of four) to fit a piecewise cubic polynomial and calculate the estimated value of the missing time point (at a uniform 0.5 second interval). For example, if is the timestamp of the known data point, For the corresponding value, interpolation calculation is required ( ), then Construct a cubic polynomial on an interval , its coefficient By ensuring that the polynomial and its first and second order derivatives are at the connection point and Continuously determine, so as to calculate As a fill value, the maximum allowed continuous interpolation length is set to 10 data points (i.e. 5 seconds). Missing segments exceeding this length are marked as unreliable data. Then, the data scale of specific physical quantities is unified. Specifically, linear normalization operations are performed on temperature values ​​(including supply water temperature and return water temperature) and flow values. The normalization formula is: ,in is the original value, is the normalized value, and The lower and upper limits of the reasonable operating range of the physical quantity are determined based on historical operating data statistics (such as the minimum and maximum values ​​of historical data) or the equipment nameplate ratings. For example, the temperature Set to 5℃, Set to 25℃, flow rate Set to 0 , Set to 1.1 times the system design maximum flow rate (for example, the design maximum flow rate is 500 ,but 550 ), so that the normalized value falls within the interval [0, 1]. Subsequently, in order to smooth the data and eliminate the instantaneous fluctuation noise introduced by the sensor or the transmission process, all parameters (including the normalized temperature and flow, as well as the original scale pump frequency, valve opening, and pressure difference) are processed using the sliding average filter method. The width of the sliding window is set to 5 consecutive data points, that is, for the current time point The value of , and its filtered value is ,For the case that there are less than 5 data points at the beginning of the sequence, the ,window size is reduced (e.g., the first two points are averaged at 2 points, the third point is averaged at 3 points, and the fourth point is averaged at 4 points).,After interpolation, normalization (for temperature and flow), and sliding average filtering, the final ,dataset contains the denoised parameter values ​​of each device at each ,calibrated timestamp (0.5 second interval), forming the system operation ,status feature set.

[0061] The steps for obtaining the input features for prediction calculation are:

[0062] Based on the system operating status feature set, we extracted time series data for supply water temperature, return water temperature, flow rate, differential pressure readings, pump frequency, and valve opening feedback. We then binned the temperature values ​​at 0.5°C intervals to calculate the frequency of occurrence, and binned the flow rate values ​​at 1m³ / h intervals to calculate the fluctuation amplitude. We then calculated the Pearson linear correlation coefficient between each bin parameter and the historical heat load values ​​to generate a set of parameter-heat load correlation coefficients.

[0063] Based on the parameter-heat load correlation coefficient set, the parameters with the absolute value of the Pearson linear correlation coefficient greater than 0.5 are screened, the time series lag relationship with the heat load value is verified, and a candidate set of key input quantities is generated;

[0064] Based on the candidate set of key input quantities, principal component analysis is performed on the supply water temperature and return water temperature values ​​to eliminate multicollinearity. The variance inflation factors of the remaining parameters are calculated, and the parameters with variance inflation factors greater than 5 are eliminated. A calculation factor matrix including the orthogonalized temperature principal components, flow values, and pressure difference operating readings is constructed to generate the prediction calculation input features.

[0065] Specifically, based on the system operation status feature set, the time series data of the six key parameters required for the study are first extracted, including the water supply temperature numerical series , return water temperature numerical sequence , flow numerical sequence , differential pressure operation reading value sequence , water pump frequency numerical sequence And the valve opening feedback value sequence At the same time, prepare a historical heat load numerical time series aligned with it The historical heat load value is based on the data of the system operation status characteristics, using the basic thermodynamic formula Calculated, where is the specific heat capacity of water (about 4.18 kJ / (kg·℃)), is the density of water (about 1000 kg / m³), ensuring The timestamps of the six extracted parameter sequences completely correspond to the timestamps of the six parameter sequences. Next, the temperature-related parameters (supply water temperature values ​​and return water temperature values) are binned, and the bin interval is set to 0.5°C. For example, if the temperature range is 5°C to 25°C, the created bins include [5.0, 5.5), [5.5, 6.0), ..., [24.5, 25.0). Then, the entire temperature time series data is traversed, and the frequency of data points in each 0.5°C interval bin (that is, the number of samples falling into this interval) is counted to obtain the frequency time series of each temperature bin. Similarly, the flow value sequence is processed. Perform binning and set the binning interval to 1 For example, if the flow rate range is 0 to 550 , then the created bins include [0, 1), [1, 2), …, [549, 550). For each flow bin, calculate the standard deviation of all flow values ​​falling into the bin to quantify the fluctuation amplitude within the flow range and obtain the fluctuation amplitude time series of each flow bin. Then, calculate the time series of each bin statistic (temperature bin frequency, flow bin fluctuation amplitude) and the original pump frequency value. , valve opening feedback value and differential pressure operating readings time series, respectively, and the historical heat load value series aligned with time Pearson linear correlation coefficient between , the calculation formula of Pearson correlation coefficient is ,in Represents a parameter (or bin statistic) at a point in time The numerical value of Represents the historical heat load at a point in time The numerical value of and are their means, is the length of the time series, and the Pearson correlation coefficient of each calculated parameter (including original parameters and bin statistics) with the historical heat load value is calculated. The values ​​are stored to form a parameter-heat load correlation coefficient set including parameter identifiers and corresponding correlation coefficient values.

[0066] Based on the parameter-heat load correlation coefficient set generated in the previous step, perform preliminary feature screening. The screening criterion is that the absolute value of the Pearson linear correlation coefficient must be greater than a preset threshold. The threshold is set to 0.5 based on engineering experience and understanding of the characteristics of the air-conditioning system. This threshold is used to select parameters that show a medium-intensity or higher linear relationship with the heat load. The specific operation is to traverse each entry in the parameter-heat load correlation coefficient set (for example: ('supply water temperature_bin_10.0-10.5_frequency', -0.6), ('flow_bin_100-101_standard deviation', 0.7), ('water pump frequency value', 0.4), ('valve opening feedback value', 0.1), ('pressure difference operating reading value', 0.55)) and check its correlation coefficient. The absolute value of ,like , the parameter identifier is temporarily retained. For example, 'water supply temperature_bin_10.0-10.5_frequency' (|r|=0.6>0.5), 'flow_bin_100-101_standard deviation' (|r|=0.7>0.5) and 'pressure difference operation reading value' (|r|=0.55>0.5) will be retained, while 'water pump frequency value' (|r|=0.4<=0.5) and 'valve opening feedback value' (|r|=0.1<=0.5) will be eliminated. For parameters that pass the correlation coefficient screening, it is necessary to further verify whether there is a reasonable time series lag relationship between them and the historical heat load values. The cross-correlation function is used for analysis to calculate the time series of each retained parameter. and historical heat load numerical time series Delayed at different times The cross-correlation coefficient under , inspection Within a preset physical reasonable range, for example, for parameters such as supply and return water temperature, flow rate, and pressure difference that directly affect the cooling exchange, the expected impact is usually immediate or slightly delayed (such as 0 to -5 minutes, where a negative value indicates that the parameter changes before the load changes). In e.g. Seconds interval, find the value that makes Peak latency ,like Falling into the preset physical reasonable range (e.g. seconds), it is considered that the timing relationship between this parameter and the heat load is reasonable. Through this verification, for example, the analysis found that the 'supply water temperature_bin_10.0-10.5_frequency' For -120 seconds, the value of 'flow_bin_100-101_standard deviation' is -10 seconds, the 'pressure differential running reading value' is -30 seconds, these are all within [-300, 0] seconds, then they all pass the verification. If a parameter (such as analyzing 'another parameter Z') is found If the peak delay of a parameter (+180 seconds, indicating that the load change is ahead of the parameter change) is unreasonable, it will be removed from the retention list. Finally, all parameters that have passed the correlation strength screening and time series lag relationship verification (here it may be necessary to integrate multiple temperature bins or flow bins that have passed the verification back to the original supply water temperature, return water temperature, and flow parameters if they show strong correlation and reasonable lag as a whole) are summarized to generate a candidate set of key input quantities.

[0067] Based on the candidate set of key input quantities generated in the previous step (for example, this set includes 'water supply temperature value' , 'Return water temperature value' , 'Flow value' , 'Differential pressure operating reading value' ), first deal with the potential multicollinearity problem between the supply water temperature and the return water temperature values, because these two temperature values ​​are usually highly correlated, use principal component analysis technology to reduce the dimension and orthogonalize, and extract the main change information of the two temperature series. The specific steps are: first, and The time series data is standardized by subtracting their respective means and dividing by their respective standard deviations to obtain the standardized series. and , then calculate the covariance matrix of the two standardized sequences, solve the eigenvalues ​​and eigenvectors of the covariance matrix, select the eigenvector corresponding to the maximum eigenvalue, project the standardized temperature data onto this main direction, and obtain the first principal component sequence ,this This is the orthogonalized temperature principal component, which represents the comprehensive trend of the supply and return water temperatures and is orthogonal to the second principal component (if calculated). Next, the degree of multicollinearity between the remaining parameters and this newly generated orthogonalized temperature principal component and between them is evaluated, considering the multicollinearity between the remaining parameters and this newly generated orthogonalized temperature principal component and between them. , flow value and differential pressure operating readings The parameter set composed of , calculates the variance inflation factor of each parameter, for the first parameters The formula for calculating its VIF value is ,in It will The coefficient of determination is calculated by linear regression with all other parameters as independent variables. 、 and The VIF value of each variable is set to 5, which is a common criterion for judging whether collinearity is serious in multivariate statistical analysis. A VIF value greater than 5 usually indicates that there is strong collinearity between the variable and other variables, which may affect the stability of subsequent model parameter estimation. The calculated VIF values ​​are checked one by one. For example, , , , since all VIF values ​​are less than or equal to 5, in this example, all parameters ( , , ) are retained. If the VIF value of a parameter is greater than 5 (for example ), then remove the parameter from the set, and finally, the orthogonalized temperature principal component obtained by PCA processing and other parameters that survive the VIF test (in this case, the flow rate value and differential pressure operating readings ) are combined together and arranged in chronological order to form a matrix. Each row of the matrix corresponds to a time point, and each column corresponds to a filtered and processed parameter. This matrix is ​​the calculation factor matrix, which generates the prediction calculation input features.

[0068] The steps to obtain the system state prediction value are:

[0069] Extract the orthogonalized temperature principal component values ​​based on the predicted calculation input features , flow value and differential pressure operating readings , perform time series processing on each parameter with a 2-hour lag and fill the missing values ​​with the moving average of the first 3 data points to generate a time series parameter matrix ;

[0070] Based on the time series parameter matrix , establish an ARIMA model, train the model parameters by minimizing the mean square error between the predicted value and the actual heat load value, and generate a trained prediction model;

[0071] Based on the trained prediction model, calculate the predicted value of the system status in the next 3 hours.

[0072] Specifically, based on the prediction calculation input features, that is, the calculation factor matrix generated in the previous step, the time series of the orthogonalized temperature principal component values ​​is extracted from it , time series of flow values And the time series of the differential pressure operating readings In order to construct the input features for predicting future heat load, time lag processing is performed on these three time series respectively, and the lag period is set to 2 hours. The selection of 2 hours lag is based on the results of cross-correlation analysis of the physical response characteristics of the air-conditioning system and historical data. The analysis shows that the current heat load has the strongest correlation with the temperature, flow and pressure difference combination state 2 hours ago. The specific implementation method of lag processing is to shift the data points of each time series by 2 hours into the future, that is, for any time point , using the original time point The value of If the sampling interval of the original data is 0.5 seconds, then 2 hours is equivalent to The time step of data points is 100 data points. After the lag operation is performed, the initial part of the sequence (the first 2 hours) will become undefined (missing values) because there is no earlier data. At the same time, the sporadic missing values ​​that may exist in the original sequence will also be brought into the lagged sequence. For these missing values ​​that appear in the lagged sequence (whether due to original missing values ​​or generated by the lag operation itself in the initial segment), a simple filling strategy is adopted: use the arithmetic mean of the three most recent valid data points before the missing value time point to fill it. The calculation formula is ,in is the time point of the lagged series The value that needs to be filled in, is Before If the first three valid data points cannot be found at the beginning of the sequence, the average of the first two or first valid data points will be tried in sequence. If there is no valid data point (such as the first few points of the sequence), it will remain undefined or use a preset default value (such as the global average of the sequence). However, it is more common to ignore these initial data segments that cannot be reliably filled in the subsequent model training to complete the training. 、 and After lag processing and missing value filling, the three processed time series 、 、 Merge by columns to construct a new matrix where each row represents a time point , each column represents a lagged input feature, which is the time series parameter matrix .

[0073] Based on the time series parameter matrix generated in the previous step , and combined with the actual heat load numerical time series aligned with it (as the target variable predicted by the model), build and train an autoregressive integrated moving average model (ARIMAX) with exogenous input. First, for the target variable The stationarity test of the training data is performed on the data of the training data. The commonly used method is the Augmented Dickey-Fuller Test (ADFTest). If the p-value of the test result is greater than the significance level (for example, set to 0.05), the series is considered non-stationary and needs to be processed by difference. Perform first-order difference and get , perform ADF test again, if the p value is less than 0.05, then the sequence after the first-order difference is considered Is stationary, determine the integration order If it is still not stable, continue to differentiate and test to determine , then determine the autoregressive order and the sliding average order , by plotting the stationary series (You may need to first remove exogenous variables through linear regression The autocorrelation function (ACF) and partial autocorrelation function (PACF) diagrams of the After truncating the order (becoming 0 or within the confidence interval), is the PACF plot After the order is truncated, a preliminary judgment is made based on the tailing and truncation characteristics of ACF and PACF and or use the information criterion method to try different value (e.g. from 0 to 5) and values ​​(e.g. from 0 to 5), construct multiple ARIMAX( ) model, and calculate the Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC) value of each model on the training set or validation set, and select the one that minimizes the AIC or BIC ( ) combination, for example, the optimal order is determined by the AIC minimization principle , then the structure of the ARIMAX (2, 1, 1) model is roughly ,in is the heat load after first-order difference, is a constant term, is the autoregressive coefficient, is the sliding mean coefficient, is an exogenous variable The coefficients of each component, is a white noise error term. The goal of model training is to find an optimal set of coefficients , so that the mean square error between the model prediction value and the actual value is minimized on the training data set. After the training is completed, the obtained model contains a certain order ( ) and the estimated coefficients, which is the trained prediction model.

[0074] Based on the prediction model trained and verified in the previous step, the prediction calculation of the heat load in the future is performed. The prediction time range is set to the next 3 hours. The prediction process is an iterative process. First, it is necessary to obtain or predict the exogenous variables in the next 3 hours. Since these are based on characteristics lagged 2 hours, the heat load is predicted for the next 3 hours. (in From 1 to , Represents the number of time steps corresponding to 3 hours) is needed The exogenous variable value at time For the part, the values ​​of these exogenous variables are known historical data, but for part, The moment is in the future, and the corresponding exogenous variable value is unknown. One way to deal with this situation is to 、 and These three original features themselves also establish simple forecasting models (for example, establish independent ARIMA models or exponential smoothing models for them respectively) to predict their future hours (because it is necessary The value of Maximum , so we need to predict ) values, and then these predicted values ​​are lagged for 2 hours to obtain the required future exogenous variable forecast values for , for example, all the (historical or predicted) exogenous variable sequences required for the next 3 hours have been obtained , the prediction steps are as follows: For the first prediction step , using known history Value, known historical error term (usually approximated by the historical residuals of the model) and exogenous variables , substitute the trained ARIMAX model equation to calculate , and then obtain the heat load prediction value through the inverse difference operation (For example ), for the second prediction step , then use the known history value, historical error term, and The predicted value calculated at (as Item) and (future error terms are usually set to 0), and exogenous variables , calculate , and then inverse difference to get , and so on, iteratively calculate the arrive (corresponding to 3 hours) all and , the final sequence This is the calculated heat load forecast value for the next 3 hours.

[0075] The steps to obtain the description of the system energy efficiency optimization problem are:

[0076] Combine the system state prediction value with the supply water temperature value, return water temperature value, flow value, and pressure difference operation reading value in the system operation state feature set, align the predicted data for the next 3 hours with the actual data at the current moment by timestamp, verify the data integrity and fill in the missing values ​​with the average of two adjacent time points to generate a joint state time series dataset;

[0077] Based on the joint state time series dataset, the total energy consumption calculation objective function is defined as a linear combination of the cube of the pump frequency value and the square of the flow value. Constraints are set, including that the difference between the supply and return water temperatures does not exceed 5°C and that the fluctuation range of the pressure differential operating reading does not exceed ±10% of the current value. This generates the objective function and constraint condition set.

[0078] Based on the objective function and the set of constraints, a linear programming model is constructed, and the optimization variables are determined to be the water pump frequency adjustment amount, the valve opening adjustment amount and the start and stop status of the chiller, which are integrated into a system energy efficiency optimization problem description including the objective function, constraints and optimization variables.

[0079] Specifically, the heat load forecast value sequence for the next 3 hours calculated in the previous step is (in is the total number of time steps corresponding to 3 hours), as well as key actual operating parameters extracted from the current real-time system operating status feature set, including the current water supply temperature value , return water temperature value , flow value and differential pressure operating readings , to integrate and prepare the data, first, the two parts of the above data are logically aligned with the timestamps, and a data structure containing future prediction information and the current baseline state is created. The structure contains a part called CurrentState, which stores the current actual value in the form of a dictionary or object. , and a section called PredictedLoad that stores a time series array of predicted heat load values ​​for the next 3 hours ,Next, the integrity verification of this integrated data structure is performed to ,check whether the value in CurrentState is valid and within the ,predefined reasonable physical range, for example, It should be between 5℃ and 15℃, for example. Should be greater than 0 , It should be between 100kPa and 600kPa, for example. If the current state data is invalid or missing, an error is recorded and the subsequent optimization process is aborted. At the same time, check whether the PredictedLoad sequence covers the entire next 3 hours ( time steps) and checks whether there are invalid values ​​(such as NaN) left over from the forecast model or previous processing in the sequence. Any internal missing values ​​found in the PredictedLoad sequence are filled using linear interpolation. If interpolation is not possible for missing values ​​at the beginning and end of the sequence, the nearest valid value is used to fill them. After verification and filling, the resulting data structure ensures that it contains a valid current system state benchmark and a complete future load forecast sequence, generating a joint state time series dataset.

[0080] Based on the joint state time series dataset generated in the previous step, we begin to define the objective function and operating constraints for energy efficiency optimization. The goal is to minimize the total energy consumption of the system (or its approximate value) in the next 3 hours and define the objective function for approximate calculation of total energy consumption. For the future Pump frequency in time steps Cube and flow The sum of the weighted linear combinations of the squares of , the mathematical expression is: ,in and Represents the future time step The operating frequency of the water pump and the total flow of the system are the results that need to be determined or indirectly affected by the optimization process, rather than directly from the joint state time series data set. The weight coefficient and It is used to balance the proportion of water pump energy consumption (approximately proportional to the cube of frequency) and other flow-related energy consumption (such as pipeline transmission loss or as a proxy for chiller load, approximately expressed by the square of flow) in the total energy consumption. These two coefficients are determined based on the energy consumption characteristics analysis and operating costs of the specific system. For example, through regression analysis of historical data or calculation based on equipment performance curves and electricity prices, it can be set and (Example value, the unit needs to be adjusted according to the actual physical dimension). Next, set the set of constraints that the system must meet. The first key constraint is the difference between the supply water temperature and the return water temperature, which requires that at each future time step , return water temperature and water supply temperature setpoint (This is one of the optimization variables) The difference does not exceed 5℃, that is This 5℃ temperature difference limit is based on the common industry setting to ensure that the evaporator has sufficient heat exchange while avoiding cooling waste. The second key constraint is the stability limit of the pipe network pressure difference, which requires that each time step in the future The differential pressure operating reading value The fluctuation range shall not exceed the current actual pressure difference value ±10% of CurrentState (from the joint state time series dataset), that is, the constraint condition is The ±10% fluctuation range is set to maintain the hydraulic balance of the pipe network and avoid drastic pressure changes that affect the terminal equipment adjustment and user comfort. The two key constraints mentioned above (temperature difference constraint and pressure difference fluctuation constraint) and other potential operating boundary conditions (such as equipment start and stop restrictions, frequency range, etc.) are summarized to generate an objective function and a set of constraints.

[0081] Based on the objective function and constraint set generated in the previous step, a specific optimization model is constructed. Although the original description mentioned "linear programming model", considering that the objective function includes the pump frequency The cubic term ( ), which is a nonlinear term, and the optimization problem usually includes discrete decisions such as the start and stop of the chiller (0 or 1 state). Therefore, the model actually constructed should be more accurately described as a mixed integer nonlinear programming problem. If the discrete decision of the start and stop of the unit is not considered, it is a nonlinear programming (NLP) problem. In this model, it is necessary to clearly define the optimization variables. These variables are the decision parameters that the optimization algorithm needs to find the optimal value, covering the next 3 hours ( arrive ) at each time step, the main optimization variables include: pump frequency setting value , its value range is usually limited by the physical properties and operation strategy of the pump, e.g. ;Water supply temperature set value , whose range is set according to comfort requirements and energy saving targets, e.g. ; Pressure difference setting value , whose range is set according to the characteristics of the pipe network, for example ; and each chiller Start and stop status , which is a binary variable, , where 1 represents open and 0 represents closed, and may also include the valve opening adjustment amount or set value As an auxiliary optimization variable, its range is , these time-varying optimization variables Integrate with the aforementioned objective function and constraints, and supplement the necessary physical model constraints (for example, With traffic , the thermodynamic equilibrium equation related to the supply and return water temperatures , and the pump characteristic curve equation that describes the relationship between pump head, flow rate, and frequency , as well as the minimum start and stop time limit of the unit, capacity limit, etc.), form a complete mathematical optimization problem statement, which includes the objective function to be minimized, the optimization variables that need to be determined and their value ranges, and all the equality and inequality constraints that must be satisfied. This is the description of the system energy efficiency optimization problem.

[0082] The steps for obtaining the hierarchical optimization control setting parameters are as follows:

[0083] Based on the objective function and constraint set in the system energy efficiency optimization problem description, the optimization range of the supply water temperature setpoint is limited to 15.0°C to 20.0°C. A list of temperature candidate values ​​is generated at 0.5°C intervals. For each candidate value, the corresponding pump frequency adjustment and return water temperature difference are calculated. Candidate values ​​that exceed the return water temperature difference limit of 5°C are eliminated to generate a set of feasible candidate values ​​for the supply water temperature.

[0084] Based on the set of feasible candidate values ​​for the water supply temperature, the pressure differential set value range of 200kPa to 500kPa is traversed to verify whether the flow value fluctuation corresponding to each candidate pressure differential value is within the allowable range of ±10%. The pressure differential candidate values ​​that meet the flow constraint and the corresponding chiller start and stop state combination codes are recorded to generate a joint feasible solution set for the pressure differential and chiller combination.

[0085] Based on the set of feasible solutions for pressure difference-unit combinations, the top 10% of optimal solutions are selected in ascending order of objective function values. The minimum start-stop interval and the maximum daily start-stop limit for the chiller start-stop state combination are verified. The water supply temperature set value, pressure difference set value, and chiller start-stop coding sequence that meets operational safety constraints are determined, and the hierarchical optimization control setting parameters are generated.

[0086] Specifically, based on the objective function and constraint set defined in the system energy efficiency optimization problem description, a hierarchical screening process is initiated to determine the optimal water supply temperature setting value. First, the water supply temperature setting value is The search range is limited to 15.0℃ to 20.0℃ at this stage (this range may be adjusted according to the specific season or building load characteristics and is different from the range in the global constraints of the model. For example, when the load is high in summer, this higher temperature range may be selected for optimization). It is discretized with a fixed interval step size of 0.5℃ to generate a list of candidate values: [15.0, 15.5, 16.0, 16.5, 17.0, 17.5, 18.0, 18.5, 19.0, 19.5, 20.0]℃. Then, for each candidate water supply temperature value in the list, , simulate its prediction time window in the next 3 hours ( arrive ) application, which requires calling a simplified system model or lookup table method to estimate the water temperature to meet the predicted heat load Required pump frequency adjustment and resulting return water temperature , the specific calculation process is: for a given and forecast load , first estimate the required flow For example, a target temperature difference (e.g., close to 5°C) can be used to preliminarily estimate the flow rate. , then according to the system resistance characteristic curve and the pressure difference setting value (for example, using the current pressure difference ) Calculate the required pump frequency , get the adjustment relative to the current frequency , then, based on the calculated flow Recalculate the return water temperature accurately , calculate the difference between the actual return water temperature and the supply water temperature set value , then, strictly check that arrive ), the calculated return water temperature difference Do they all satisfy the previously set constraints, i.e. This 5℃ threshold is set according to the efficient operation range of the refrigeration equipment and the pipe network design specifications. The candidate water supply temperature value exceeds this limit at any time. All of them are considered infeasible and removed from the candidate list, and the candidate values ​​that are not removed are collected to generate a set of feasible candidate values ​​for water supply temperature.

[0087] Based on the set of feasible candidate values ​​of water supply temperature generated in the previous step, further explore the feasible pressure difference setting value and chiller start and stop state combination that match it. First, determine the pressure difference setting value The traversal range is 200kPa to 500kPa. This range is determined according to the minimum and maximum pressure differences that the pipe network design can support, and a traversal step size is set, such as 10kPa, to generate a list of candidate pressure difference values: [200, 210, 220, ..., 490, 500]kPa. At the same time, all possible combinations of chiller start and stop states are generated. For example, if the system has two chillers, the possible combinations are: unit 1 on / unit 2 off (code 01), unit 1 off / unit 2 on (code 10), unit 1 on / unit 2 on (code 11), and a unique combination code is assigned to each combination. Next, a nested loop is executed: the outer loop traverses each of the feasible candidate value sets for the water supply temperature , the middle layer loops through each of the candidate pressure value lists The inner loop traverses all valid chiller start and stop status combination codes , for each formed ( , , ) triples, and evaluates their feasibility within the next 3-hour forecast window. Two main aspects need to be verified: First, the selected unit combination Is the total cooling capacity at all time steps arrive are greater than or equal to the corresponding predicted heat load , if the cooling capacity is insufficient at any time, the triplet is not feasible. Secondly, it is verified that under this setting ( , , ), the traffic generated by system operation (Need to be obtained by solving the coupled model of hydraulic balance and thermal balance) relative to the current actual flow Whether the fluctuation of (from the joint state time series data set) is within the allowable range, which is set to ±10% of the current flow, that is, checking whether For all This ±10% flow fluctuation limit is intended to maintain the relative stability of the system hydraulic conditions and avoid impact on the terminal equipment. For each triplet that passes the cooling capacity verification and flow fluctuation verification, the corresponding pressure difference candidate value is recorded. Combined coding of chiller start and stop status , and compare them with the current feasible values ​​of water supply temperature Associate, and calculate and record the objective function value corresponding to the triple , all recorded feasible combinations and their objective function values ​​are summarized to generate a set of joint feasible solutions of pressure difference-unit combination.

[0088] Based on the pressure difference-unit combination joint feasible solution set generated in the previous step, this set includes all solutions that meet the basic operating constraints (temperature difference, flow fluctuation, cooling capacity matching) ( , , , ) combination, perform the final screening and verification process, first, according to the objective function value corresponding to each feasible solution (representing the expected approximate total energy consumption), sort all the solutions in the set in ascending order, and then select the top-ranked optimal solution subset from the sorted list, and set the selection ratio to the top 10%. The top 10% rather than only the optimal solution is selected in order to further consider the more stringent safety and equipment protection constraints in actual operation in the approximate energy consumption optimal solution set. If there are a total of feasible solutions, then choose For example, if there are 200 feasible solutions, the first 20 are selected. Next, the first 10% of the optimal solutions are checked for detailed operational safety constraints, with the focus on checking the start and stop state combinations of the chillers. Whether the operating restrictions of the equipment will be violated during actual execution, specifically including: checking whether the start-stop actions implied by the combination (relative to the change in the current operating state) meet the minimum start-up interval time of each chiller (for example, it takes at least 15 minutes to start again after shutdown) and the minimum operating time (for example, it takes at least 30 minutes to run after startup before it can be shut down). These time limits are specified by the equipment manufacturer. Secondly, it is necessary to combine historical operating records (such as the number of starts and stops in the past 24 hours) to check whether the implementation of the combination will cause the cumulative number of starts of any unit in a single day to exceed the preset upper limit (for example, each unit is started a maximum of 8 times a day). This restriction is also to protect the equipment and extend its life. Only when the start-stop state combination of the chiller corresponding to a solution fully meets all the above-mentioned minimum start-stop interval time and maximum start-stop number restrictions, the solution is considered to be finally executable. From all the top 10% solutions that pass the final safety check, the one with the lowest objective function value is selected. The optimal solution is obtained, and the water supply temperature set value, pressure difference set value and chiller start and stop code sequence that meets the operation safety constraints (which may be a detailed start and stop plan covering the next 3 hours, or a combined code applied at the starting time) are determined to generate the hierarchical optimization control setting parameters.

[0089] The steps to obtain the equipment adjustment difference calculation results are as follows:

[0090] Extract the water supply temperature set value, pressure difference set value and chiller start and stop code from the hierarchical optimization control setting parameters, and simultaneously obtain the corresponding real-time water supply temperature, real-time pressure difference and water pump frequency in the system operation status feature set. Calculate the accumulated temperature deviation and pressure difference change rate at three consecutive time points according to the timestamp to generate a dynamic deviation parameter set.

[0091] Based on the dynamic deviation parameter set, the allowable boundary threshold of the pump frequency adjustment is calculated using the following formula:

[0092] ;

[0093] in, is the boundary threshold, is the reference proportional coefficient of the pump frequency and temperature deviation, For the The temperature change rate at a time point, The difference between the current pressure difference set value and the real-time value, is the pressure difference reference constant, Number the current time point;

[0094] Based on the boundary threshold, the difference between the target frequency adjustment amount and the real-time frequency is limited to Within the range, the adjustment amount that exceeds the threshold is truncated according to the boundary value, and the device adjustment difference calculation result is generated, which includes the timestamp, the truncated adjustment amount and the constraint logic.

[0095] Specifically, extract the hierarchical optimization control setting parameters determined in the previous step, specifically including the target water supply temperature setting value for the next control cycle , target pressure difference setting value And the start and stop status combination code of the chiller These set values ​​are the recommended operating points calculated based on energy efficiency optimization and operation constraints. At the same time, by accessing the system operation status feature set, the current moment Latest real-time operating data, including real-time water supply temperature , real-time pressure difference And the real-time operating frequency of the water pump Next, use these set values ​​and real-time values ​​to calculate the key indicators that reflect the current dynamic deviation of the system. First, calculate three consecutive time points (the current moment , the previous moment , a moment before ) of the water supply temperature deviation, i.e. for (Need to get from historical data or cache and ), and then calculate the cumulative amount of these deviations, for example simply summing: This cumulative amount reflects the degree of continuous deviation of recent temperature control. Secondly, to calculate the rate of change of real-time pressure difference, the real-time pressure difference value at the current and past two time points is required. , , (also obtained from the historical records of the system operation status feature set), calculate the pressure difference change at adjacent time points and , and then estimate the recent average rate of change, which reflects the stability of the recent network hydraulic conditions. The calculated temperature deviation accumulation and pressure difference change rate, together with the original data points used to calculate them and the current timestamp, are used to calculate the average rate of change. , organized together to generate a dynamic deviation parameter set.

[0096] formula: The formula is useful in that it is used to calculate the dynamic boundary threshold of the pump frequency adjustment amount. It is not a fixed limit, but a dynamic adjustment of the allowable control action amplitude according to the real-time dynamic characteristics of the system (temperature change rate, pressure difference deviation). Its design reflects the balance consideration of the stability and responsiveness of the control system.

[0097] The steps to obtain the parameters are: Represents the base proportional coefficient of the pump frequency and temperature deviation, for example, calculated from 15 sets of qualified test data The average value is 1.8Hz / ℃, so set , this value reflects the average sensitivity of the pump frequency to the supply water temperature of this particular water system.

[0098] The steps to obtain the parameters are: Represents a discrete time point The rate of change of water supply temperature, which reflects the instantaneous speed of temperature change, for example, at time point , get , , , , then calculate: °C / second, °C / second, °C / sec.

[0099] The steps to obtain the parameters are: Indicates the current time point , the difference between the network pressure difference setting value and the real-time measurement value. For example, the target pressure difference set in the current control cycle is kPa, the current real-time pressure difference read from the sensor is kPa, then the pressure difference deviation kPa.

[0100] The steps to obtain the parameters are: Is a pressure difference reference constant used to adjust the pressure difference deviation in the inverse tangent function Scale it so that it is close to the boundary threshold The contribution of the air conditioning water system falls into an appropriate range. The setting method can be: refer to the design documents of the air conditioning water system to obtain the pressure difference value of the supply and return water main under the design working conditions; or analyze the historical data of the long-term operation of the system (for example, retrieve the data of the past three months). Data), calculate the average or median pressure difference during normal operation (excluding start-up and shutdown or large load change periods), and use this statistical value as For example, by analyzing historical data, it is found that the pressure difference of the system is stable between 280kPa and 320kPa for 70% of the time, and the median is about 300kPa. kPa.

[0101] The steps to obtain the parameters are: The number or index representing the time point at which the current calculation occurs is a counter for discrete-time systems. For example, if the control period is 0.5 seconds, then The value is incremented by 1 every 0.5 seconds, representing the current data frame being processed or the moment of control calculation.

[0102] Substituting the parameters into the formula, we get 0.0034, which shows that at the current moment The boundary threshold of the allowable pump frequency adjustment is calculated based on the temperature change stability and pressure difference deviation of the system. It is about 0.0034Hz. This value is very small, which means that under the current system state (small temperature change and small pressure difference deviation), the allowable frequency adjustment range is very limited, and almost no adjustment is allowed. This helps maintain the stability of the system and prevent unnecessary control disturbances when approaching the target state. If the calculation result Larger, e.g. Hz, it means that the system deviates far from the set point or changes rapidly, and a maximum frequency adjustment of ±0.5Hz is allowed in the next control step. The value is used directly in the next step to limit the originally calculated frequency adjustment.

[0103] The time-varying boundary threshold calculated in the previous step , to perform the final constraint processing on the control adjustment of the pump frequency, first, it is necessary to obtain an unconstrained "target frequency adjustment" , this raw adjustment is calculated by the main control algorithm (such as PID controller, fuzzy controller or other optimization adjustment strategy based on the current system deviation and set value), and it represents the frequency change value expected to be applied in a single cycle without considering dynamic constraints, for example, a frequency change based on the water supply temperature deviation The PID controller may calculate that the current frequency needs to be increased by 0.1Hz, that is, Hz, then, this original adjustment and the calculated dynamic boundary threshold Compare and determine whether it is within the allowed range The specific logic is: if The absolute value of is less than or equal to (Right now ), it means that the adjustment amount is considered safe and appropriate under the current working conditions and no modification is required. The frequency adjustment amount finally adopted It is equivalent to , for example, if Hz and Hz, then Hz, the constraint logic is recorded as "untriggered constraint", otherwise, if If it exceeds this range, it needs to be truncated: , indicating that the expected adjustment amount is too large (positive), then the final adjustment amount is forced to be limited to the upper boundary value, that is, , for example, if Hz and Hz, then Hz, the record constraint logic is "exceeded the upper boundary, truncated", similarly, if , indicating that the expected adjustment amount is too large (negative), then the final adjustment amount is forced to be limited to the lower boundary value, that is, , for example, if Hz and Hz, then Hz, the record constraint logic is "exceeded the lower boundary, truncated", after this comparison and possible truncation processing, the obtained This is to determine the frequency adjustment value to be sent to the water pump (relative to the current frequency ), the current timestamp , this final truncated adjustment The calculation results of the equipment adjustment difference are generated by combining the applied constraint logic (not triggered, exceeded the upper boundary, exceeded the lower boundary).

[0104] The steps to obtain the real-time device adjustment instruction sequence are:

[0105] Extract the timestamp adjustment amount sequence from the equipment adjustment difference calculation results, calculate the absolute value mean and standard deviation of each adjustment amount according to the time window, eliminate abnormal adjustments that exceed 3 times the standard deviation range, and generate a cleaned adjustment amount sequence;

[0106] Based on the adjustment amount sequence after cleaning, the delay factor of the control instruction is calculated using the following formula:

[0107] ;

[0108] in, is the delay factor, is the end index of the current time window, For the The amount of adjustment after cleaning at each time point, is the device response delay base, is the instruction decay time constant;

[0109] Based on the delay effectiveness factor, delay compensation is added to each adjustment amount, and the compensated instruction is checked to see whether it exceeds the maximum step change rate of the pump frequency, generating a real-time equipment adjustment instruction sequence.

[0110] Specifically, extract the timestamp and corresponding final frequency adjustment value (the value after the dynamic boundary threshold constraint) contained in the equipment adjustment difference calculation result generated in the previous step. , forming a time series ,Next, in order to identify and process the possible abnormal fluctuations or noise ,spikes in the sequence, a statistical method based on time ,window is used to perform data cleaning, setting a fixed width time window, for example, ,window width The window size is set to 60 data points, which corresponds to a time length of 30 seconds (if the data sampling interval is 0.5 seconds). This window size selection is intended to cover enough data points for effective statistical analysis while being able to adapt quickly to changes in system conditions. For each time point in the sequence , after examining (or including within) data points (if insufficient data If all available data points are used), all adjustments within the time window are calculated. The arithmetic mean of the absolute values ​​of and standard deviation Then, the 3-sigma criterion is applied to determine the last point (or current point) in the window. ) Whether it is an abnormal value, the judgment condition is: if , then the point is considered as an abnormal adjustment. The 3 Sigma criterion is a commonly used anomaly detection method in statistics. It is based on the fact that if the data roughly obeys the normal distribution, about 99.7% of the data points should fall within the range of the mean plus or minus 3 times the standard deviation. Values ​​outside this range can be considered as low-probability events or abnormal points. For the detected abnormal adjustment, a replacement strategy is adopted instead of direct deletion. Its value is modified to the value determined by the statistical boundary, that is, the new adjustment If the original adjustment is not judged as abnormal, its value remains unchanged. By applying this sliding window-based statistical calculation and 3 sigma criterion outlier detection and processing to the entire adjustment sequence, a cleaned adjustment sequence with smoother data fluctuations and less noise interference is generated. .

[0111] formula: The formula is useful in that it is used to calculate the delay factor of the control instruction. Its core purpose is to estimate the cumulative effect or "momentum" of the system due to actuator or process delays based on the control adjustments (after cleaning) at the most recent time points and their time decay characteristics. The calculation results can be used to compensate for future control instructions and consider the delay effect in advance, thereby improving the response speed and accuracy of the control system and reducing overshoot or oscillation.

[0112] The steps to obtain the parameters are: Represents the discrete time index or number corresponding to the end point of the time window used for the current calculation. For example, in a control system with a period of 0.5 seconds, if the system has been running for 100 seconds, the current Maybe the first time steps.

[0113] The steps to obtain the parameters are: It means in time points (where The range is arrive ) The pump frequency adjustment after cleaning, for example, at time , found from the cleaned sequence Hz, Hz, and Hz.

[0114] The steps to obtain the parameters are: It is the base of the device response delay, which reflects the inherent delay or inertia of the controlled object (water pump and VFD system) in responding to the control command. This parameter is usually obtained through experimental calibration. For example, through testing, it is found that under the current operating conditions, single-step adjustment exceeding 0.5Hz is likely to cause oscillation. In this case, you can try to set Hz, as the basis for the adjustment amplitude.

[0115] The steps to obtain the parameters are: Is the instruction decay time constant, which determines the speed at which the historical adjustment weight decays over time. For example, if the system sampling period is 0.5 seconds, and it is expected that the influence of the adjustment amount will be significantly attenuated within 1.5 seconds (3 sampling points), then you can set (Unit: number of sampling cycles / number of steps).

[0116] Substituting the parameters into the formula, the result is 0.331, which shows that: at the current time point Calculated delay factor It is approximately 0.331. This value is positive, indicating that the overall trend of the weighted adjustment at the last three time points is positive (increasing frequency). Its size of 0.331 is relative to the base number. In terms of , it represents a medium-intensity positive adjustment "momentum". This factor will be used in the next step to calculate the specific delay compensation amount, and a positive compensation value proportional to the momentum will be added to the subsequent control instructions to offset part of the system response delay.

[0117] Based on the delay factor calculated in the previous step , the adjustment amount after cleaning Perform delay compensation processing and perform the final safety check. First, calculate the specific delay compensation value, which is equal to the delay effectiveness factor Multiply by a preset compensation gain coefficient to set a compensation coefficient ,For example Hz, this coefficient needs to be adjusted and optimized according to the actual system delay characteristics and control effect, the compensation value , add this compensation value to the post-cleaning adjustment amount at the current time point to obtain the initial compensation adjustment amount , for example, if and Hz, Hz, then the compensation value is Hz, the adjustment amount after compensation is Hz, then the pump operation safety check is performed on this compensated adjustment, mainly to check whether it will cause the pump frequency to change too quickly and exceed the maximum frequency step change rate (SlewRate) allowed by the equipment. A maximum allowable step change rate needs to be set. , which is determined by the technical specifications or protection parameter settings of the water pump or variable frequency drive (VFD), such as setting Hz / second, the adjustment amount after compensation Converted into the rate of change per unit time, that is ,in is to control the time step (for example 0.5 seconds), and then compare Is it greater than ,if , indicating that the adjustment rate is within the allowable range, and the final control instruction adjustment amount It is equivalent to ,if , it means that the adjustment amount after compensation is too drastic and needs to be limited to limit the final control instruction adjustment amount to the maximum allowable range of variation, that is, , for example, if Hz, s, then the rate is Hz / s, less than Hz / s, so Hz, the current timestamp The final adjustment after delay compensation and maximum step change rate calibration is Combined to generate the current instruction in the real-time device adjustment instruction sequence.

[0118] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A deep learning-based air conditioning chilled water control method, characterized in that: The following steps are involved: Collect the status of the air-conditioning chiller, including pump frequency and valve opening feedback, as well as the pipe network supply water temperature, return water temperature, flow rate and pressure differential operation readings, synchronize and structure the signals, and establish a system operation status feature set; Based on the system operation state feature set, key input quantities that affect heat load changes are identified to obtain prediction calculation input features, and based on the prediction calculation input features, prediction calculations are performed to obtain the performance status of the chiller water pump to obtain a system state prediction value; Combining the system state prediction value with the current system operating state feature set, a calculation framework with the goal of minimizing total energy consumption is constructed and associated operating boundaries are set to obtain a description of the system energy efficiency optimization problem. Based on the description of the system energy efficiency optimization problem, a solution is performed to determine the water supply temperature, pressure difference set value and chiller operation combination, and establish hierarchical optimization control setting parameters; Comparing the target set values ​​in the hierarchical optimization control setting parameters with the real-time operating values ​​fed back by the system operating status feature set, calculating the required adjustment amount of the water pump frequency, and obtaining a device adjustment difference calculation result; based on the device adjustment difference calculation result, generating control instructions according to the device response characteristics and safety constraints, and organizing them in a time sequence to obtain a real-time device adjustment instruction sequence; The steps for obtaining the prediction calculation input features are: Based on the system operating status feature set, extract the time series data of the supply water temperature, return water temperature, flow rate, differential pressure operation reading, pump frequency, and valve opening feedback value. Statistically analyze the frequency of occurrence of the temperature value at 0.5°C intervals and the fluctuation amplitude of the flow rate value at 1m³ / h intervals. Calculate the Pearson linear correlation coefficient between each bin parameter and the historical heat load value to generate a parameter-heat load correlation coefficient set. Based on the parameter-heat load correlation coefficient set, parameters with an absolute value of the Pearson linear correlation coefficient greater than 0.5 are screened, the time series lag relationship with the heat load value is verified, and a candidate set of key input quantities is generated; Based on the candidate set of key input quantities, principal component analysis is performed on the supply water temperature values ​​and the return water temperature values ​​to eliminate multicollinearity, variance inflation factors of the remaining parameters are calculated, parameters with variance inflation factors greater than 5 are eliminated, and a calculation factor matrix including orthogonalized temperature principal components, flow values, and pressure difference operating reading values ​​is constructed to generate prediction calculation input features; The steps for obtaining the system state prediction value are: Based on the predicted calculation input features, the orthogonalized temperature principal component values ​​are extracted , flow value and differential pressure operating readings , perform time series processing on each parameter with a 2-hour lag and fill the missing values ​​with the moving average of the first 3 data points to generate a time series parameter matrix ; Based on the time series parameter matrix , establish an ARIMA model, train the model parameters by minimizing the mean square error between the predicted value and the actual heat load value, and generate a trained prediction model; Based on the trained prediction model, the system state prediction value within the next 3 hours is calculated.

2. The air conditioning chilled water control method based on deep learning according to claim 1, characterized in that: The steps for obtaining the system operation status feature set are: Collect the pump frequency value, valve opening feedback value, pipe network supply water temperature value, return water temperature value, flow value and pressure difference operation reading value of the air conditioning chiller. All values ​​under the same device code are aligned to the system global clock reference at the time of collection. After marking the device code and timestamp, the original operating parameter set is generated; Based on the original operating parameter set, for values ​​with missing timestamps or clock offsets between devices, the cubic spline interpolation method is used to complete the data sequence at 0.5-second time intervals. The temperature and flow values ​​are linearly normalized respectively. The sliding average filtering method with a window width of 5 data points is used to eliminate instantaneous fluctuation noise, and a system operating status feature set containing device codes, correction timestamps and denoising parameters is generated.

3. The air conditioning chilled water control method based on deep learning according to claim 1, characterized in that: The steps for obtaining the description of the system energy efficiency optimization problem are: Combine the system state prediction value with the supply water temperature value, return water temperature value, flow value and pressure difference operation reading value in the system operation state feature set, align the predicted data for the next 3 hours with the actual data at the current moment by timestamp, verify the data integrity and fill in the missing values ​​with the average of two adjacent time points to generate a joint state time series data set; Based on the joint state time series data set, the total energy consumption calculation objective function is defined as a linear combination of the cube of the water pump frequency value and the square of the flow value. Constraints are set, including that the difference between the supply water temperature and the return water temperature does not exceed 5°C and the fluctuation range of the pressure difference operating reading does not exceed ±10% of the current value. The objective function and constraint condition set are generated; Based on the objective function and the set of constraints, a linear programming model is constructed, and the optimization variables are determined to be the water pump frequency adjustment amount, the valve opening adjustment amount and the start and stop status of the chiller, which are integrated into a system energy efficiency optimization problem description including the objective function, constraints and optimization variables.

4. The air conditioning chilled water control method based on deep learning according to claim 1, characterized in that: The steps for obtaining the hierarchical optimization control setting parameters are as follows: Based on the objective function and constraint set in the system energy efficiency optimization problem description, the optimization range of the supply water temperature set point is limited to 15.0°C to 20.0°C. A list of temperature candidate values ​​is generated at 0.5°C intervals. For each candidate value, the corresponding water pump frequency adjustment amount and the return water temperature difference are calculated. Candidate values ​​that exceed the return water temperature difference limit of 5°C are eliminated to generate a set of feasible candidate values ​​for the supply water temperature. Based on the set of feasible candidate values ​​for the water supply temperature, traverse the pressure differential set value range of 200 kPa to 500 kPa, verify whether the flow value fluctuation corresponding to each candidate pressure differential value is within the allowable range of ±10%, record the candidate pressure differential values ​​that meet the flow constraint and the corresponding chiller start and stop state combination codes, and generate a joint feasible solution set of pressure differential-unit combinations; Based on the set of feasible solutions for the pressure difference-unit combination, the top 10% of the optimal solutions are selected in ascending order of the objective function value, the minimum start-stop interval time and the maximum daily start-stop limit of the chiller start-stop state combination are verified, the water supply temperature set value, the pressure difference set value and the chiller start-stop code sequence that meets the operation safety constraints are determined, and the hierarchical optimization control setting parameters are generated.

5. The air conditioning chilled water control method based on deep learning according to claim 1, characterized in that: The steps for obtaining the equipment adjustment difference calculation result are: Extract the water supply temperature setting value, pressure difference setting value and chiller start / stop code from the hierarchical optimization control setting parameters, synchronously obtain the corresponding real-time water supply temperature, real-time pressure difference and water pump frequency in the system operation status feature set, calculate the temperature deviation accumulation and pressure difference change rate at three consecutive time points according to the timestamp, and generate a dynamic deviation parameter set; Based on the dynamic deviation parameter set, the allowable boundary threshold of the water pump frequency adjustment amount is calculated, and the calculation formula is: ； in, is the boundary threshold, is the reference proportional coefficient of the pump frequency and temperature deviation, For the The temperature change rate at a time point, The difference between the current pressure difference set value and the real-time value, is the pressure difference reference constant, Number the current time point; Based on the boundary threshold, the difference between the target frequency adjustment amount and the real-time frequency is limited to Within the range, the adjustment amount that exceeds the threshold is truncated according to the boundary value, and the device adjustment difference calculation result is generated, which includes the timestamp, the truncated adjustment amount and the constraint logic.

6. The air conditioning chilled water control method based on deep learning according to claim 1, characterized in that: The steps for obtaining the real-time device adjustment instruction sequence are: Extracting a timestamp adjustment amount sequence from the equipment adjustment difference calculation result, calculating the absolute value mean and standard deviation of each adjustment amount according to the time window, eliminating abnormal adjustments that exceed a range of 3 times the standard deviation, and generating a cleaned adjustment amount sequence; Based on the adjustment amount sequence after cleaning, the delay effect factor of the control instruction is calculated, and the calculation formula is: ； in, is the delay factor, is the end index of the current time window, For the Adjust the amount after cleaning at each time point. is the device response delay base, is the instruction decay time constant; Based on the delay effectiveness factor, delay compensation is added to each adjustment amount, and it is checked whether the compensated instruction exceeds the maximum step change rate of the water pump frequency, and a real-time equipment adjustment instruction sequence is generated.

7. The air conditioning chilled water control system according to any one of claims 1 to 6, characterized in that: include: The status data acquisition module collects the status of the chiller, obtains water pump frequency feedback, valve opening feedback, pipe network supply water temperature reading, return water temperature reading, flow reading and pressure difference operation reading, performs signal synchronization processing, performs signal structure processing, and establishes synchronized structured operation reading; The operation feature extraction module establishes a system operation status feature set based on the synchronous structured operation readings, identifies the key input quantities that affect the heat load change, obtains the input features required for the prediction calculation, and establishes the system operation status feature set; The unit performance prediction module selects prediction calculation input features based on the system operating status feature set, performs prediction operations, obtains the performance status of the chiller pump, evaluates the expected interaction between the supply water temperature setting and the cooling tower fan speed adjustment under different loads, determines the expected system pressure drop of the evaporator and condenser, and obtains the system status prediction value; The energy efficiency optimization setting module combines the system state prediction value with the current system operation state feature set to build a calculation framework with the goal of minimizing total energy consumption. It sets the associated operation boundaries, considers the capacity limit of the chiller and the pressure requirement of the pipe network, solves the system energy efficiency optimization problem description, evaluates different water supply temperature set points, pressure difference set points and chiller operation combination schemes, determines the optimal water supply temperature set point, pressure difference set point and chiller operation combination, and establishes the hierarchical optimization control setting parameters. The equipment instruction generation module compares the target setting values ​​in the hierarchical optimization control setting parameters with the real-time operating values ​​fed back by the system operation status feature set, calculates the required adjustment amount of the water pump frequency, and obtains the equipment adjustment difference calculation result. Based on the equipment adjustment difference calculation result, the control instruction is generated according to the equipment response characteristics to obtain a real-time equipment adjustment instruction sequence.

Images