Water chilling unit model updating method based on incremental learning

Abstract

The invention discloses a water chilling unit model updating method based on incremental learning, and the method comprises the steps: carrying out the incremental training of an original water chilling unit model through the local parameter updating and knowledge retention mechanism of an incremental learning algorithm, continuously employing the latest operation data of a water chilling unit at a current stage, and dynamically adjusting the model parameters, the progressive offset of the equipment performance is synchronously adapted, and meanwhile, the problem of disastrous forgetting that the performance of a new model on old tasks is obviously reduced when the new data is adopted to update the model is prevented. The method is easy to expand to modeling of water chilling units of different types and models, solves the problem that a previous model prediction value deviates from an actual operation value due to equipment aging or environment change, and has a good actual engineering application prospect.

Description

Technical Field

[0001] The present invention belongs to multiple fields such as HVAC energy-saving technology and machine learning, and specifically relates to a chiller model updating method based on incremental learning. Background Art

[0002] Chillers consume a significant portion of the energy in HVAC systems. To predict chiller performance under different operating conditions, analyze their operating patterns, and optimize their energy-efficient operation, chiller modeling is necessary to facilitate unified management of equipment in the computer room. Traditional modeling methods typically construct static parameter models based on historical equipment operating data.

[0003] However, in actual engineering applications, it has been found that the phenomenon of gradual performance drift of chillers is common in long-term operation: the attenuation of heat transfer efficiency caused by scaling of the heat exchanger, the performance degradation caused by mechanical wear of the compressor, the temperature sensitivity changes of refrigerant physical parameters with the seasons and other time-varying factors will cause systematic deviations between the actual operating characteristic curve of the equipment and the original model. This performance drift presents a nonlinear cumulative effect in engineering practice, causing the prediction accuracy of the static model based on initial working condition training to continue to decline over time. Therefore, developing a model update method with dynamic adaptability to achieve the synchronous evolution of model parameters and equipment performance drift has become a key technical challenge to improve the applicability of chiller engineering. Summary of the Invention

[0004] To solve the above problems, the present invention discloses a chiller model updating method based on incremental learning, which has dynamic adaptability and realizes the synchronous evolution of model parameters and equipment performance drift, so as to more accurately predict the operating conditions of the chiller in actual engineering projects in the long term.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A chiller model updating method based on incremental learning includes the following steps:

[0007] S1: Acquisition of on-site chiller model;

[0008] S2: Collect the latest operating data of the chiller at this stage;

[0009] S3: pre-processing the latest collected chiller operation data;

[0010] S4: The original model of the chiller is trained and updated using the latest operating data of the chiller that has been preprocessed in S3 through an incremental learning algorithm to obtain a new model that can accurately reflect the characteristics of the chiller at this stage.

[0011] Furthermore, the step S1 includes:

[0012] The collected chiller model formula is unified as follows:

[0013] COP=a1×T e +a2×T c +a3×Q+a4×Q 2 +a5×T e ×Q+a6×T c ×Q+a7;

[0014] Among them, a1, a2…a7 are the parameters to be fitted, T e is the chilled water outlet temperature of the chiller, T c is the cooling water inlet temperature of the chiller, and Q is the cooling capacity of the chiller.

[0015] Furthermore, the step S2 includes:

[0016] Collect all operating data of the on-site chiller during operation within the specified time range, including start and stop status, chilled water inlet and outlet water temperature, cooling water inlet and outlet water temperature, load rate, cooling capacity and power data.

[0017] Furthermore, the step S3 includes

[0018] S31: Calculate the COP of the chiller using the collected operating data, and calculate the chiller cooling efficiency (COP) using the formula COP=Q / W, where Q is the chiller cooling capacity and W is the chiller power;

[0019] S32: converting non-numeric operating parameter data of the chiller equipment into numeric values, such as the equipment switch status, where 1 represents on and 0 represents off;

[0020] S33: Time tag standardization: Standardize the time tags of the data. The normalized time is determined by the following formula: t = [t0 + (n-1) × δ], (n = 1, 2, 3, ...), where δ is the data step size and t0 is the first data time;

[0021] S34: Abnormal value elimination and correction: Abnormal values are eliminated and corrected based on the actual possible range of the operating parameters; for example, the collected temperature exceeds the actual possible operating temperature range of the chiller;

[0022] S35: Missing value filling: For missing signal parameters, the missing values are filled using the data of the previous and next time points using interpolation or extrapolation.

[0023] S36: Data standardization: Perform standardization on all obtained parameters:

[0024] xnormalized =(x-μ) / σ

[0025] Where x is the specific parameter, μ is the mean value of all parameters, and σ is the standard deviation of the parameter.

[0026] Get the standardized parameter set.

[0027] Furthermore, the step S4 includes:

[0028] S41: Input the collected original model of the chiller, which is in the form of:

[0029] COP=a1×T e +a2×T c +a3×Q+a4×Q 2 +a5×T e ×Q+a6×T c ×Q+a7;

[0030] S42: input the chiller operation data set after the normalization process obtained in S3;

[0031] S43: Calculate the total gradient J(θ); including the loss gradient and the regularization gradient αR(θ)

[0032] S44: Calculate the weight vector:

[0033]

[0034] Where L is the loss function; R(θ) is the regularization term; α is the regularization strength coefficient; θ is the weight vector; η is the learning rate; is the gradient of the loss function L with respect to θ; f(x i ,θ) is the model’s predicted output for the input sample; y i is the true label of the sample; is the gradient of the regularization function R(θ) with respect to θ.

[0035] S45: Iteratively calculate and update the weight vector with a fixed step size η until convergence;

[0036] S46: The weight vector at the time of convergence is the weight vector value of the updated model to form the final prediction model.

[0037] The beneficial effects of the present invention are:

[0038] 1. By introducing an incremental learning mechanism to dynamically update the chiller model and fine-tuning the model's local parameters through real-time collection of the latest operating data, the model can track progressive offset characteristics such as heat exchanger efficiency decay and compressor performance degradation, reducing prediction errors and increasing the prediction accuracy of key operating parameters (such as cooling capacity and COP) during long-term operation.

[0039] 2. The incremental training strategy only requires storing the current window data, eliminating the need to retain the entire historical data set for a long time. This reduces storage resource usage while avoiding the large amount of computing power required for traditional full reconstruction, ensuring the continuity of the system's prediction function.

[0040] This method systematically solves the engineering contradiction between model inaccuracy and high maintenance costs in the long-term operation of chillers through a lightweight and continuous model update strategy, providing high-precision, low-resource consumption technical support for the digital operation and maintenance of refrigeration systems, and has good prospects for practical engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a detailed flow chart of the present invention;

[0042] Figure 2 A detailed flow chart for preprocessing the collected operating data;

[0043] Figure 3 Updated the case flow chart for a specific model based on incremental learning. DETAILED DESCRIPTION

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0045] As shown in the figure, the chiller model updating method based on incremental learning of the present invention includes:

[0046] S1: Acquisition of on-site chiller model;

[0047] S2: Collect the latest operating data of the chiller at this stage;

[0048] S3: pre-processing the collected chiller operation data;

[0049] S4: The original model of the chiller is trained and updated using the latest operating data of the chiller that has been preprocessed in S3 through an incremental learning algorithm to obtain a new model that can accurately reflect the characteristics of the chiller at this stage.

[0050] As a specific embodiment, step S1 includes:

[0051] The collected chiller model formula is unified as follows:

[0052] COP=a1×T e +a2×T c +a3×Q+a4×Q 2 +a5×T e ×Q+a6×Tc ×Q+a7;

[0053] Among them, a1, a2…a7 are the parameters to be fitted, T e is the chilled water outlet temperature of the chiller, T c is the cooling water inlet temperature of the chiller, and Q is the cooling capacity of the chiller.

[0054] The step S2 comprises:

[0055] Collect all operating data of the on-site chiller during operation within the specified time range, including start and stop status, chilled water inlet and outlet water temperature, cooling water inlet and outlet water temperature, load rate, cooling capacity and power data.

[0056] The step S3 is as follows Figure 2 As shown, including:

[0057] S31: Calculate the COP of the chiller using the collected operating data, and calculate the chiller cooling efficiency (COP) using the formula COP=Q / W, where Q is the chiller cooling capacity and W is the chiller power;

[0058] S32: converting non-numeric operating parameter data of the chiller equipment into numeric values, such as the equipment switch status, where 1 represents on and 0 represents off;

[0059] S33: Time tag standardization: Standardize the time tags of the data. The normalized time is determined by the following formula: t = [t0 + (n-1) × δ], (n = 1, 2, 3, ...), where δ is the data step size and t0 is the first data time;

[0060] S34: Abnormal value elimination and correction: Abnormal values are eliminated and corrected based on the actual possible range of the operating parameters; for example, the collected temperature exceeds the actual possible operating temperature range of the chiller;

[0061] S35: Missing value filling: For missing signal parameters, the missing values are filled using the data of the previous and next time points using interpolation or extrapolation.

[0062] S36: Data standardization: Perform standardization on all obtained parameters:

[0063] x normalized =(x-μ) / σ

[0064] Where x is the specific parameter, μ is the mean value of all parameters, and σ is the standard deviation of the parameter.

[0065] Get the standardized parameter set.

[0066] The step S4 is as follows Figure 3 Shown include:

[0067] Specifically, the stochastic gradient descent algorithm SGDClassifier in incremental learning is selected to update the chiller model.

[0068] S41: Input the collected original model of the chiller, which is in the form of:

[0069] COP=a1×T e +a2×T c +a3×Q+a4×Q 2 +a5×T e ×Q+a6×T c ×Q+a7;

[0070] S42: input the chiller operation data set after the normalization process obtained in S3;

[0071] S43: Calculate the total gradient J(θ); including the loss gradient and the regularization gradient αR(θ)

[0072] S44: Calculate the weight vector:

[0073]

[0074] Where L is the loss function; R(θ) is the regularization term; α is the regularization strength coefficient; θ is the weight vector; η is the learning rate; is the gradient of the loss function L with respect to θ; f(x i ,θ) is the model’s predicted output for the input sample; y i is the true label of the sample; is the gradient of the regularization function R(θ) with respect to θ.

[0075] S45: Iteratively calculate and update the weight vector with a fixed step size η until convergence;

[0076] S46: The weight vector at the time of convergence is the weight vector value of the updated model to form the final prediction model.

[0077] SGDClassifier is a linear classifier based on stochastic gradient descent in the Scikit-learn library. It updates the weights of model parameters by performing gradient descent (SGD) on a new dataset without using historical datasets to calculate gradients.

[0078] In Python, you can directly use the SGDClassifier function in the scikit-learn library to update the model using the latest data. The penalty term alpha in the model is the regularization term coefficient. The larger this value is, the greater the penalty for the parameters in the model. In the specific implementation, the penalty term alpha is set to 0.005.

[0079] It should be noted that the above content merely illustrates the technical idea of the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. A chiller model updating method based on incremental learning, characterized in that: The steps include: S1: Acquisition of on-site chiller model; S2: Collect the latest operating data of the chiller at this stage; S3: pre-processing the collected chiller operation data; S4: The original model of the chiller is trained and updated using the latest operating data of the chiller that has been preprocessed in S3 through an incremental learning algorithm to obtain a new model that can accurately reflect the characteristics of the chiller at this stage.

2. The chiller model updating method based on incremental learning according to claim 1, characterized in that: The acquisition site chiller model described in step S1, The chiller model formula is: <h2 style=";text-align:left;direction:ltr">COP = a1×T<h2 style=";text-align:left;direction:ltr"> e <h2 style=";text-align:left;direction:ltr"> +a2×T<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> +a3×Q+a4×Q×Q+a5×T<h2 style=";text-align:left;direction:ltr"> e <h2 style=";text-align:left;direction:ltr"> ×Q+a6×T<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> ×Q+a7; Among them, a1, a2…a7 are the parameters to be fitted, T e is the chilled water outlet temperature of the chiller, T c is the cooling water inlet temperature of the chiller, and Q is the cooling capacity of the chiller.

3. The chiller model updating method based on incremental learning according to claim 1, characterized in that: Step S2 is to collect the latest operating data of the chiller. Collect all operating data of the on-site chiller during operation within the specified time range, including start and stop status, chilled water inlet and outlet water temperature, cooling water inlet and outlet water temperature, load rate, cooling capacity and power data.

4. The chiller model updating method based on incremental learning according to claim 1, characterized in that: Preprocessing the collected refrigeration machine operation data as described in step S3; S31: Calculate the COP of the chiller using the collected operating data, and calculate the chiller refrigeration efficiency using the formula COP=Q / W, where Q is the chiller cooling capacity and W is the chiller power; S32: converting non-numerical operating parameter data of the refrigeration equipment into numerical values; S33: Time tag standardization: Standardize the time tags of the data. The normalized time is determined by the following formula: t = [t0 + (n-1) × δ], (n = 1, 2, 3, ...), where δ is the data step size and t0 is the first data time; S34: Outlier elimination and correction: Eliminate and correct outliers based on the actual possible range of operating parameters; S35: Missing value filling: For missing signal parameters, use the data of the previous and next time points to fill the missing values using interpolation or extrapolation. S36: Data standardization: Perform standardization on all obtained parameters: x normalized =(x-μ) / σ Where x is the specific parameter, μ is the average value of all parameters, and σ is the standard deviation of the parameter. Get the standardized parameter set.

5. The chiller model updating method based on incremental learning according to claim 1, characterized in that: The incremental learning model updating method described in step S4 is characterized by: S41: Input the collected original model of the chiller, which is in the form of: <h2 style=";text-align:left;direction:ltr">COP = a1×T<h2 style=";text-align:left;direction:ltr"> e <h2 style=";text-align:left;direction:ltr"> +a2×T<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> +a3×Q+a4×Q<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> +a5×T<h2 style=";text-align:left;direction:ltr"> e <h2 style=";text-align:left;direction:ltr"> ×Q+a6×T<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> ×Q+a7; S42: input the chiller operation data set after the normalization process obtained in S3; S43: Calculate the total gradient J(θ); including the loss gradient and the regularization gradient αR(θ); S44: Calculate the weight vector: Where L is the loss function; R(θ) is the regularization term; α is the regularization strength coefficient; θ is the weight vector; η is the learning rate; is the gradient of the loss function L with respect to θ; f(x i ,θ) is the model’s predicted output for the input sample; y i is the true label of the sample; is the gradient of the regularization function R(θ) with respect to θ, S45: Iteratively calculate and update the weight vector with a fixed step size η until the loss function converges; S46: The weight vector at the time of convergence is the weight vector value of the updated model to form the final prediction model.

Images