Central air conditioner water chilling unit energy efficiency ratio prediction method based on XGBoost and LSTM, medium and equipment
By employing a hybrid modeling approach combining XGBoost and LSTM, and integrating simulation and real data, the problem of data scarcity and adaptability in chiller energy efficiency prediction was solved, achieving high-precision and stable energy efficiency prediction and supporting multiple rounds of iterative optimization throughout the chiller's lifecycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG DIOR TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-01
AI Technical Summary
Existing energy efficiency prediction models for chillers face challenges such as the scarcity and limited distribution of real data, discrepancies between simulation and real data distribution, difficulty in balancing learning efficiency and adaptability, and a lack of adaptive mechanisms, resulting in insufficient prediction accuracy and stability.
A hybrid modeling approach based on XGBoost and LSTM is adopted. The initial training is dominated by simulation data, and dynamic weight adjustment and incremental learning with real data are combined to achieve the fusion of simulation and real data, thereby improving the model's adaptability and generalization performance.
It achieves high-precision and stable prediction of the energy efficiency ratio of chiller units, supports long-term deployment and iterative optimization, reduces maintenance costs, and is suitable for continuous learning in industrial environments.
Smart Images
Figure CN121960143A_ABST
Abstract
Description
A method, medium, and equipment for predicting the energy efficiency ratio of central air conditioning chillers based on XGBoost and LSTM. Technical Field
[0001] This invention relates to the field of chiller technology, and in particular to a method, medium, and equipment for predicting the energy efficiency ratio of central air conditioning chillers based on XGBoost and LSTM. Background Technology
[0002] As a core component of industrial refrigeration systems, chiller units' operating energy efficiency (often expressed as COP, Coefficient of Performance) significantly impacts the overall energy system efficiency and operating costs. With the development of intelligent technologies, more and more companies hope to leverage machine learning or deep learning models to accurately predict the operating energy efficiency of chiller units, thereby achieving energy efficiency optimization and predictive maintenance. However, the following challenges commonly exist in the process of chiller unit operation modeling:
[0003] 1. Scarce and limited distribution of real data: In the early deployment phase, the actual data collected from the operation of chiller units is limited and the coverage of scenarios is insufficient, which can easily lead to underfitting of the model.
[0004] 2. There are differences in the distribution between simulation data and real data: Although simulation data can be generated in batches, there is a certain deviation from the actual operating state, and the generalization ability of the model obtained by direct training is limited; 3. It is difficult for the model to balance learning efficiency and real adaptability: Relying solely on real data for training is slow and the model is unstable; relying entirely on simulation data will sacrifice accuracy; 4. Lack of adaptive mechanism for new data: Existing prediction models are often trained in one go, lacking the mechanism for later fine-tuning and incremental updates, making it difficult to adapt to the dynamic changes in the operating state of the equipment.
[0005] Therefore, there is an urgent need for a hybrid modeling algorithm that integrates the advantages of simulation and real data, can dynamically adapt to changes in data distribution, and supports continuous optimization, so as to achieve high-precision and stable prediction of the energy efficiency ratio of chiller units. Summary of the Invention
[0006] This invention addresses the problems of existing technologies by providing a method, medium, and equipment for predicting the energy efficiency ratio of central air conditioning chiller units based on XGBoost and LSTM. It integrates modeling capabilities using simulation and real data, effectively utilizing the coverage and physical constraints of simulation data while combining the reliability of real data to achieve efficient modeling. By progressively increasing weights from real data, it achieves natural transitional learning, significantly enhancing the model's adaptability and generalization performance. By freezing some model parameters and performing local fine-tuning, it improves the long-term update efficiency and deployment convenience of the model, reducing maintenance costs. Combining XGBoost's ability to model static features with LSTM's ability to capture temporal patterns, it achieves structure-temporal feature fusion. It is suitable for long-term deployment and iterative optimization in industrial environments, supporting multiple iterations and continuous learning of the model throughout the chiller unit's lifecycle.
[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0008] This invention provides a method for predicting the energy efficiency ratio of central air conditioning chiller units based on XGBoost and LSTM, which includes the following steps:
[0009] Step S1: Initial training phase dominated by simulation data:
[0010] Step S11, Data Preparation and Generation: Generate simulation data covering different working conditions, environmental parameters and equipment configurations;
[0011] Step S12, Feature Processing:
[0012] The temporal features from the simulation data are input into the LSTM sub-model. The LSTM sub-model then processes these features using a bidirectional LSTM, concatenating the forward and backward hidden state sequences along the feature dimension to obtain the output sequence of the bidirectional LSTM. Finally, an attention mechanism layer focuses on key time segments within the temporal sequence, forming the temporal feature representation vector H. t ;
[0013] Non-temporal features from the simulation data are input into the Xgboost sub-model. The Xgboost sub-model learns the non-linear relationship between non-temporal features and COP through training, and outputs a structural feature representation vector X. xgb ;
[0014] Step S13: Fuse the temporal feature representation vector H t and structural feature representation vector X xgb The fusion weight vector g is obtained, and the fusion vector Z is generated. The fusion vector Z is input to the output layer for prediction, and the continuous predicted value of the chiller unit COP is output through the output layer.
[0015] Step S2, Dynamic Weight Adjustment Stage: As the chiller unit is put into operation and begins to accumulate real data;
[0016] Step S22, Dynamic Sample Weighted Training:
[0017] During the actual operation of the chiller unit, real operating data is continuously collected, and the corresponding real COP value is calculated. The same preprocessing process as the simulation data is performed on the real data.
[0018] Step S23, Hybrid Training and Adaptive Weight Adjustment:
[0019] Construct a hybrid training set: containing both simulated and real data;
[0020] A dynamic weighting mechanism is introduced: initially, simulation data is given high weights, and as the amount of real data increases, the weight of real data in the loss function is gradually increased;
[0021] Retrain the hybrid model to smoothly transition to the distribution of real-world operating conditions;
[0022] Step S3, Incremental Learning and Model Update Phase: Continuously monitor the model's prediction error on new real data. When the error exceeds the threshold or enough new data is accumulated, trigger the incremental learning mechanism.
[0023] The simulation data includes environmental features, static equipment features, dynamic operating features, and periodic time codes.
[0024] Wherein, the temporal feature representation vector H output by the LSTM sub-model t ,
[0025] Among them, X t Select data from the first N operating times of the chiller unit.
[0026] Wherein, the fusion weight vector g, g = σ(W) g ·[X xgb H t ]+b g ); where, [X xgb H t [] represents two vectors representing the concatenated structure and time sequence; Wg is a learned weight matrix; σ is the Sigmoid function, which is used to adjust X using the Sigmoid function and the learnable weight matrix. xgb With H t The concatenated vectors are used to calculate the fusion weight vector g, which has values between 0 and 1.
[0027] Wherein, the fusion vector Z = g·X xgb +(1-g)·H t .
[0028] In step S13, the output layer is a fully connected regression layer adapted for chiller unit COP prediction, and its input is a fusion vector Z generated by a gated fusion mechanism. This output layer is configured with a learnable weight matrix Wout that matches the dimension of the fusion vector Z, and a bias term bout. After inputting the fusion vector Z into the output layer, the COP prediction value is obtained through a linear transformation operation. The specific calculation formula is as follows:
[0029] COP pred =W out ·Z+b out .
[0030] In step S2, real operating data is continuously collected during the actual operation of the chiller unit, and the corresponding real COP value is calculated.
[0031] As the chiller units come into operation and real data gradually accumulates, the training loss weight ratio between simulation and real samples is controlled:
[0032] Weighting coefficients:
[0033] w real (t)=min(1,α·F days (t)·Q(t)·R(t));
[0034] w sim ((t)=1-w real (t);
[0035] Wherein, ωreal(t) refers to the weight system of the real data at time t, and the value range of ωreal(t) is [0,1], which is used to characterize the contribution ratio of the real data in the current training.
[0036] α refers to the weight growth ratio coefficient, which is used to control the growth rate of the weights in the actual data.
[0037] Fdays(t) refers to the cumulative number of days of real data at time t, which directly represents the amount of real working condition data accumulated.
[0038] Q(t) refers to the validity coefficient of the real data at time t, which is used to filter high-quality real data;
[0039] R(t) refers to the distribution similarity coefficient between real and simulated data at time t, which is obtained by calculating the KL divergence of the feature distributions of real and simulated data.
[0040] ωsim(t) refers to the weight coefficient of the simulation data at time t, which is derived from ωreal(t). It ensures that the sum of the weights of the two types of data is 1, thus achieving a smooth transition of weights.
[0041] Total training loss function:
[0042]
[0043] Lsim refers to the loss on simulated data, Lreal refers to the loss on real data, and ωreal and ωsim are weight coefficients that change dynamically with the number of training rounds or the cumulative amount of real data.
[0044] In step S2, the specific process of hybrid training and adaptive weight adjustment is as follows:
[0045] Initial Connection: After step S1, the chiller unit is put into actual operation and collects real operating data, and enters the dynamic weight adjustment stage. When t=0 initially, wreal=0 and wsim=1.
[0046] Weight update trigger: For every preset duration of accumulated real data, a weight update is triggered to calculate F days(t), Q(t), and R(t) at the current time t.
[0047] Weight calculation: The weights of the real data are obtained through the weight coefficient calculation formula, and the weights of the simulated data are also obtained.
[0048] Training iteration: Calculate the total loss using the total loss function, backpropagate the loss using optimization algorithms such as Adam, and update the model parameters.
[0049] Stable Termination: When wreal increases to 1, the training enters a state of "realism-driven, simulation-constrained" training.
[0050] The specific method for the incremental learning and model update stage in step S3 is as follows:
[0051] Step S31: Establish a model performance monitoring mechanism: Calculate the model's prediction metrics on new real data daily / hourly: MAE, RMSE; Set a first trigger condition and a second trigger condition for the incremental learning and model update phase. If the first trigger condition or / and the second trigger condition are met, the incremental learning trigger mechanism will be triggered.
[0052] The first triggering condition is defined as the cumulative real data volume N. real If the proportion exceeds 60%, the second triggering condition is a decrease in model performance. Compare the MAE and RMSE metrics on the validation sets before and after the test. If the condition is met...
[0053] RMSE new -RMSE old ≥δ perf If the percentage is ≥10%, the incremental learning mechanism is triggered; where RMSE newThe root mean square error of the current model on the latest real data validation set is used to characterize the model's current COP prediction performance for the latest chiller unit operating conditions.
[0054] RMSE old The root mean square error (RMSE) on the corresponding validation set after the model's last training or incremental learning is used as the baseline value for model performance; δpref is a preset performance degradation threshold. When the increment of RMSE reaches or exceeds this threshold, it is determined that the model performance has degraded, thus satisfying the second triggering condition for incremental learning.
[0055] Step S32, Freeze and defrost configuration of model layers:
[0056] Freeze the XGBoost structure branches;
[0057] Freeze the LSTM embedding layer;
[0058] Unfreeze the Attention layer;
[0059] Unfreeze gated fusion layer: This layer can adjust the fusion weights of "structural features and temporal features" to adapt to changes in feature dependencies under new data;
[0060] Unfreeze the output layer: Allows for fine-tuning of its parameters to match the COP prediction output pattern of the latest real data;
[0061] S33. Preparation of incremental training data:
[0062] A training set is constructed from newly added real data, while a validation set is constructed by selecting independent, latest real data.
[0063] S34, Low learning rate warm start training;
[0064] The loss function is "COP prediction mean square error (MSE) of the latest real data". The loss is backpropagated only to the unfrozen layers to update their parameters. The RMSE of the validation set is continuously monitored during training. Fine-tuning is terminated when the performance of the validation set no longer improves for several consecutive rounds.
[0065] S35. Model Update and Recording:
[0066] The fine-tuned thawing layer parameters are integrated with the frozen layer parameters to obtain an updated hybrid model, which is used for subsequent chiller COP prediction. At the same time, the RMSE of the current model on the validation set is recorded as the RMSEold benchmark for the next incremental learning.
[0067] The present invention also provides a computer storage medium storing computer instructions, which, when invoked, are used to execute the aforementioned method for predicting the energy efficiency ratio of central air conditioning chiller units based on XGBoost and LSTM.
[0068] The present invention also provides an electronic device comprising: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the aforementioned XGBoost and LSTM-based method for predicting the energy efficiency ratio of a central air conditioning chiller.
[0069] The beneficial effects of this invention are:
[0070] This invention integrates modeling capabilities with simulation and real data: it effectively utilizes the coverage and physical constraints of simulation data, combined with the reliability of real data, to achieve efficient modeling; it achieves natural transitional learning by gradually increasing weights from real data, significantly enhancing the model's adaptability and generalization performance; it improves long-term model update efficiency and deployment convenience by freezing some model parameters and performing local fine-tuning, reducing maintenance costs; it combines XGBoost's ability to model static features with LSTM's ability to capture temporal patterns, achieving structure-temporal feature fusion; and it is suitable for long-term deployment and iterative optimization in industrial environments: supporting multiple iterations and continuous learning of the model throughout the lifecycle of chiller units. Attached Figure Description
[0071] Figure 1 is a flowchart of the initial training phase dominated by simulation data in this invention.
[0072] Figure 2 is a flowchart of the dynamic weight adjustment stage of the present invention.
[0073] Figure 3 is a flowchart of the incremental learning and model update stages of the present invention. Detailed Implementation
[0074] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention. The present invention will be described in detail below with reference to the accompanying drawings.
[0075] Example 1
[0076] In Embodiment 1 of this application, as shown in Figures 1 to 3, a method for predicting the energy efficiency ratio of a chiller unit based on a hybrid model of XGBoost and LSTM includes the following steps:
[0077] Step S1: Initial training phase dominated by simulation data:
[0078] Step S11, Data Preparation and Generation: Generate simulation data covering different working conditions, environmental parameters and equipment configurations;
[0079] Step S12, Feature Processing:
[0080] The temporal features from the simulation data are input into the LSTM sub-model. The LSTM sub-model then processes these features using a bidirectional LSTM, concatenating the forward and backward hidden state sequences along the feature dimension to obtain the output sequence of the bidirectional LSTM. Finally, an attention mechanism layer focuses on key time segments within the temporal sequence, forming the temporal feature representation vector H. t ;
[0081] Non-temporal features from the simulation data are input into the Xgboost sub-model. The Xgboost sub-model learns the non-linear relationship between non-temporal features and COP through training, and outputs a structural feature representation vector X. xgb ;
[0082] Step S13: Fuse the temporal feature representation vector H t and structural feature representation vector X xgb The fusion weight vector g is obtained, and the fusion vector Z is generated. The fusion vector Z is input to the output layer for prediction, and the continuous predicted value of the chiller unit COP is output through the output layer.
[0083] Step S2, Dynamic Weight Adjustment Stage: As the chiller unit is put into operation and begins to accumulate real data;
[0084] Step S22, Dynamic Sample Weighted Training:
[0085] During the actual operation of the chiller unit, real operating data is continuously collected, and the corresponding real COP value is calculated. The same preprocessing process as the simulation data is performed on the real data.
[0086] Step S23, Hybrid Training and Adaptive Weight Adjustment:
[0087] Construct a hybrid training set: containing both simulated and real data;
[0088] A dynamic weighting mechanism is introduced: initially, simulation data is given high weights, and as the amount of real data increases, the weight of real data in the loss function is gradually increased;
[0089] Retrain the hybrid model to smoothly transition to the distribution of real-world operating conditions;
[0090] Step S3, Incremental Learning and Model Update Phase: Continuously monitor the model's prediction error on new real data. When the error exceeds the threshold or enough new data is accumulated, trigger the incremental learning mechanism.
[0091] In this embodiment, the simulation data includes environmental features, equipment static features, operational dynamic features, and periodic time encoding. Specifically, this embodiment uses a large amount of physically constrained simulation data to generate training samples, covering multi-dimensional scenarios such as different operating conditions, environmental parameters, and equipment configurations; the constructed features include:
[0092] Environmental characteristics: temperature and humidity, seasonal indicators, and meteorological information;
[0093] Static characteristics of the equipment: model, rated power, refrigerant type;
[0094] Operating dynamics: load factor, condensing / evaporating temperature series, historical COP;
[0095] Periodic time coding (such as sine and cosine coding); input standardization, missing data imputation, and anomaly detection.
[0096] In this embodiment, the simulation data sources are: 1. Historical weather data (temperature, humidity, and meteorological data). 2. Static equipment characteristics (device static characteristics) obtained from the manufacturer's nameplate information or product manual. 3. Operational characteristics are derived by fitting a formula based on the machine's performance model or sample data provided by the manufacturer, and then covering possible operating conditions to generate simulation data.
[0097] In this embodiment, the Xgboost sub-model takes non-time-series features (such as unit structural parameters, static operating condition parameters, etc.) as input, and learns the non-linear correlation between these features and COP through training; after training, the sub-model outputs a structural feature representation vector X. xgb It is a high-dimensional abstract representation of the input structured features, which can effectively characterize the nonlinear influence of structured and non-temporal features on COP.
[0098] In this embodiment of the application, in the LSTM sub-model, X t The time series input for constructing the LSTM sub-model is the data from the previous N operating times of the chiller unit, including operating status parameters (evaporation temperature, condensation temperature, flow rate) and historical loads; LSTM(X t ): Bidirectional LSTM for X t The processing output; a bidirectional LSTM consists of a "forward LSTM" and a "backward LSTM", with the forward LSTM processing X at time steps t = 1 - T. t The forward hidden state sequence is obtained; the backward LSTM processes X at time steps t = 1 - T. t The backward hidden state sequence is obtained, and the two are concatenated along the feature dimension to obtain the output sequence of the bidirectional LSTM; where Attention(LSTM(X)) t () is the attention mechanism layer, used to focus on key time segments in a time sequence.
[0099] In this embodiment of the application, the temporal feature representation vector H output by the LSTM sub-model t , Among them, X t Select data from the first N operating times of the chiller unit.
[0100] In this embodiment of the application, the fusion weight vector g,
[0101] g=σ(W g ·[X xgb H t ]+b g ); where, [X xgb H t [] represents two vectors representing the concatenated structure and time sequence; Wg is a learned weight matrix; σ is the Sigmoid function, which is used to adjust X using the Sigmoid function and the learnable weight matrix. xgb With H t The concatenated vectors are used to calculate the fusion weight vector g, which has values between 0 and 1.
[0102] Wherein, the fusion vector Z = g·X xgb +(1-g)·H t .
[0103] In step S13, the output layer is a fully connected regression layer adapted for chiller unit COP prediction, and its input is a fusion vector Z generated by a gated fusion mechanism. This output layer is configured with a learnable weight matrix Wout that matches the dimension of the fusion vector Z, and a bias term bout. After inputting the fusion vector Z into the output layer, the COP prediction value is obtained through a linear transformation operation. The specific calculation formula is as follows:
[0104] COP pred =W out ·Z+b out ;
[0105] Specifically, the weight matrix W of the output layer out With bias term b out It is learned gradually through a phased model training process; in the initial training phase, which mainly uses simulation data, the weight matrix W is first trained. out With bias term b out Perform random initialization (e.g., using a normal or uniform distribution to initialize parameter values); dynamic weight adjustment: adaptive parameter optimization, as real data gradually accumulates and the model enters the dynamic weight adjustment phase, W... out With b out The parameters from step S1 will be iteratively updated (through backpropagation of the loss function);
[0106] Since COP is a continuous energy efficiency index, the output layer uses a linear activation function to directly output the result of the above linear transformation as the final predicted COP value of the chiller unit. This predicted value also serves as the basis for loss calculation during the model training phase (initial simulation data training and dynamic weight adjustment of real data), as well as the target reference for model performance evaluation and high-level parameter fine-tuning during the incremental learning phase.
[0107] In this embodiment of the application, in step S2, real operating data is continuously collected during the actual operation of the chiller unit, and the corresponding real COP value is calculated.
[0108] As the chiller units come into operation and real data gradually accumulates, the training loss weight ratio between simulation and real samples is controlled:
[0109] Weighting coefficients:
[0110] w real (t)=min(1,α·F days (t)·Q(t)·R(t));
[0111] w sim (t)=1-w real (t);
[0112] Wherein, ωreal(t) refers to the weight system of the real data at time t, and the value range of ωreal(t) is [0,1], which is used to characterize the contribution ratio of the real data in the current training.
[0113] α refers to the weight growth ratio coefficient, which is used to control the growth rate of the weights in the actual data.
[0114] Fdays(t) refers to the cumulative number of days of real data at time t, which directly represents the amount of real working condition data accumulated.
[0115] Q(t) refers to the validity coefficient of the real data at time t, which is used to filter high-quality real data;
[0116] R(t) refers to the distribution similarity coefficient between real and simulated data at time t, which is obtained by calculating the KL divergence of the feature distributions of real and simulated data.
[0117] ωsim(t) refers to the weight coefficient of the simulation data at time t, which is derived from ωreal(t). It ensures that the sum of the weights of the two types of data is 1, thus achieving a smooth transition of weights.
[0118] Total training loss function:
[0119]
[0120] Lsim refers to the loss on simulated data, Lreal refers to the loss on real data, and ωreal and ωsim are weight coefficients that change dynamically with the number of training rounds or the cumulative amount of real data.
[0121] In step S2, the specific process of hybrid training and adaptive weight adjustment is as follows:
[0122] Initial Connection: After step S1, the chiller unit is put into actual operation and collects real operating data, and enters the dynamic weight adjustment stage. When the initial t=0, wreal=0 and wsim=1, the model parameters of step S1 are reused.
[0123] Weight update trigger: Every time a preset period of real data is accumulated (e.g., daily), a weight update is triggered to calculate F days(t) (cumulative days), Q(t) (data validity), and R(t) (distribution similarity) at the current time t.
[0124] Weight calculation: The weights of the real data are obtained through the weight coefficient calculation formula, and the weights of the simulated data are also obtained.
[0125] Training iteration: Calculate the total loss using the total loss function, backpropagate the loss using optimization algorithms such as Adam, and update the model parameters.
[0126] Stable Termination: When wreal increases to 1, the training enters a state of "realism-driven, simulation-constrained" training.
[0127] The specific method for the incremental learning and model update stage in step S3 is as follows:
[0128] Step S31: Establish a model performance monitoring mechanism: Calculate the model's prediction metrics on new real data daily / hourly: MAE, RMSE; Set a first trigger condition and a second trigger condition for the incremental learning and model update phase. If the first trigger condition or / and the second trigger condition are met, the incremental learning trigger mechanism will be triggered.
[0129] The first triggering condition is defined as the cumulative real data volume N. real If the proportion exceeds 60%, the second triggering condition is a decrease in model performance. Compare the MAE and RMSE metrics on the validation sets before and after the test. If the condition is met...
[0130] RMSE new -RMSE old ≥δ perf If the percentage is ≥10%, the incremental learning mechanism is triggered; where RMSE new The root mean square error of the current model on the latest real data validation set is used to characterize the model's current COP prediction performance for the latest chiller unit operating conditions.
[0131] RMSE old The root mean square error (RMSE) on the corresponding validation set after the model's last training or incremental learning is used as the baseline value for model performance; δpref is a preset performance degradation threshold. When the increment of RMSE reaches or exceeds this threshold, it is determined that the model performance has degraded, thus satisfying the second triggering condition for incremental learning.
[0132] Step S32, Freeze and defrost configuration of model layers:
[0133] Freeze XGBoost structure branches: maintain its ability to learn the "non-linear correlation between structured, non-temporal features and COP", and avoid fine-tuning from destroying the feature extraction logic accumulated in the early stage;
[0134] Freeze the LSTM embedding layer: retain its original time series modeling capability and maintain the effect of capturing the time series dependence of the historical operating conditions of the chiller unit;
[0135] Unfreeze the Attention layer: This allows the layer to refocus on key time segments in the latest real data (such as periods of abnormal unit operation);
[0136] Unfreeze gated fusion layer: This layer can adjust the fusion weights of "structural features (XGBoost output) and temporal features (LSTM output)" to adapt to changes in feature dependencies under new data;
[0137] Unfreeze the output layer: Allows for fine-tuning of its parameters to match the COP prediction output pattern of the latest real data;
[0138] S33. Preparation of incremental training data:
[0139] The training set is constructed from the newly added real data using a "small batch random sampling" method (to avoid overfitting caused by full training), while a validation set is constructed using independent, latest real data (for real-time monitoring of fine-tuning effects).
[0140] S34, Low learning rate warm start training; Use a "low learning rate" (such as 1 / 10 of the learning rate in the initial training phase) for training to avoid significantly perturbing the already stable model parameters;
[0141] Using the mean squared error (MSE) of the latest real data prediction as the loss function, the loss is backpropagated only to the unfrozen layers (Attention layer, gated fusion layer, and output layer) to update their parameters. During training, the RMSE of the validation set is continuously monitored. Fine-tuning is terminated when the validation set performance no longer improves for several consecutive rounds (or reaches the preset number of iterations).
[0142] S35. Model Update and Recording:
[0143] The fine-tuned thawing layer parameters are integrated with the frozen layer parameters to obtain an updated hybrid model, which is used for subsequent chiller COP prediction. At the same time, the RMSE of the current model on the validation set is recorded as the RMSEold benchmark for the next incremental learning.
[0144] Specifically, compared with existing methods, the embodiments of this application have the following significant advantages:
[0145] Modeling capabilities that integrate simulation and real data: Effectively utilize the coverage and physical constraint characteristics of simulation data, combined with the reliability of real data, to achieve efficient modeling;
[0146] Dynamic weight adjustment mechanism: By gradually increasing weights through real data, natural transition learning is achieved, which significantly enhances the model's adaptability and generalization performance;
[0147] It has incremental learning capabilities: by freezing some model parameters and making local fine-tuning, it improves the efficiency of long-term model updates and the convenience of deployment, and reduces maintenance costs;
[0148] Enhancing model robustness: Combining XGBoost's ability to model static features with LSTM's ability to capture temporal patterns, structural-temporal feature fusion is achieved;
[0149] Suitable for long-term deployment and iterative optimization in industrial environments: Supports multiple iterations and continuous learning of the model throughout the lifecycle of the chiller unit.
[0150] Example 2
[0151] Embodiment 2 of this application also provides a computer storage medium storing computer instructions. When the computer instructions are invoked, they are used to execute the aforementioned method for predicting the energy efficiency ratio of central air conditioning chiller units based on XGBoost and LSTM.
[0152] Example 3
[0153] Embodiment 3 of this application also provides an electronic device, wherein the electronic device includes: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to execute the aforementioned XGBoost and LSTM-based central air conditioning chiller unit energy efficiency ratio prediction method.
[0154] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the present invention without departing from the scope of the present invention are within the scope of the present invention.
Claims
1. A method for predicting the energy efficiency ratio of central air conditioning chiller units based on XGBoost and LSTM, characterized in that, Includes the following steps: Step S1, Initial Training Phase Dominated by Simulation Data: Step S11, Data Preparation and Generation: Generate simulation data covering different working conditions, environmental parameters, and equipment configurations; Step S12, Feature Processing: Input the temporal features from the simulation data into the LSTM sub-model. The LSTM sub-model uses a bidirectional LSTM to process the data, concatenating the forward and backward hidden state sequences along the feature dimension to obtain the output sequence of the bidirectional LSTM. Then, an attention mechanism layer focuses on key time segments in the temporal sequence to form the temporal feature representation vector H. t The non-temporal features from the simulation data are input into the Xgboost sub-model. The Xgboost sub-model learns the non-linear relationship between the non-temporal features and COP through training, and outputs a structural feature representation vector X. xgb Step S13: Fuse the temporal feature representation vector H t and structural feature representation vector X xgb The fusion weight vector g is obtained, and the fusion vector Z is generated; the fusion vector Z is input to the output layer for prediction, and the continuous predicted value of the chiller unit COP is output through the output layer; Step S2, dynamic weight adjustment stage: as the chiller unit is put into operation and begins to accumulate real data; Step S22, Dynamic Sample Weighted Training: Continuously collect real operating data during the actual operation of the chiller unit, calculate the corresponding real COP value, and perform the same preprocessing process on the real data as on the simulation data; Step S23, Hybrid Training and Adaptive Weight Adjustment: Construct a hybrid training set containing both simulated and real data; Introduce a dynamic weighting mechanism: Initially assign high weights to simulated data, and gradually increase the weights of real data in the loss function as the amount of real data increases; Retrain the hybrid model to smoothly transition to the distribution of real working conditions; Step S3, Incremental Learning and Model Update Stage: Continuously monitor the prediction error of the model on new real data, and trigger the incremental learning mechanism when the error exceeds the threshold or when enough new data is accumulated.
2. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that: The simulation data includes environmental features, static equipment features, dynamic operating features, and periodic time codes.
3. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that: The temporal feature representation vector H output by the LSTM sub-model t , Among them, X t Select data from the first N operating times of the chiller unit.
4. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that: The fusion weight vector g, g = σ(W) g ·[X xgb H t ]+b g ); where, [X xgb H t [] represents two vectors representing the concatenated structure and time sequence; Wg is a learned weight matrix; σ is the Sigmoid function, which is used to adjust X using the Sigmoid function and the learnable weight matrix. xgb With H t The concatenated vectors are used to calculate the fusion weight vector g, which takes values between 0 and 1; where the fusion vector Z = g·X xgb +(1-g)·H t .
5. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that: In step S13, the output layer is a fully connected regression layer adapted for chiller unit COP prediction, and its input is the fusion vector Z generated by the gated fusion mechanism. This output layer is configured with a learnable weight matrix Wout that matches the dimension of the fusion vector Z, and a bias term bout. After inputting the fusion vector Z into the output layer, the COP prediction value is obtained through linear transformation. The specific calculation formula is: COP pred =W out ·Z+b out .
6. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that: In step S2, real operating data is continuously collected during the actual operation of the chiller unit, and the corresponding real COP value is calculated. As the chiller units come online and real data gradually accumulates, the training loss weight ratio between simulation and real samples is controlled: weight coefficient: w real (t)=min(1,α·F days (t)·Q(t)·R(t);w sim (t)=1-w real (t); where ωreal(t) refers to the weight system of the real data at time t, and the value range of ωreal(t) is [0,1], which is used to characterize the contribution ratio of the real data in the current training; α refers to the weight growth ratio coefficient, which is used to control the growth rate of the weight of the real data; Fdays(t) refers to the cumulative number of days of real data at time t, directly representing the amount of real working condition data accumulated; Q(t) refers to the validity coefficient of real data at time t, used to screen high-quality real data; R(t) refers to the distribution similarity coefficient of real-simulation data at time t, obtained by calculating the KL divergence of the feature distributions of real and simulation data; ωsim(t) refers to the weight coefficient of simulation data at time t, derived from ωreal(t), ensuring that the sum of the weights of the two types of data is 1, achieving a smooth transition of weights; Training total loss function: Lsim refers to the loss on simulated data, Lreal refers to the loss on real data, and ωreal and ωsim are weight coefficients that change dynamically with the number of training rounds or the cumulative amount of real data.
7. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that, In step S2, the specific process of hybrid training and adaptive weight adjustment is as follows: Initial connection: After step S1, the chiller unit is put into actual operation and collects real operating data, and enters the dynamic weight adjustment stage. When t=0 initially, wreal=0 and wsim=1; Weight update trigger: Every time real data is accumulated for a preset time, a weight update is triggered, and Fdays(t), Q(t), and R(t) at the current time t are calculated. Weight Calculation: The weights of the real data are obtained through the weight coefficient calculation formula, and the weights of the simulated data are also obtained. Training Iteration: The total loss is calculated using the total loss function, and the model parameters are updated by backpropagating the loss through optimization algorithms such as Adam. Stable Termination: When wreal increases to 1, the training enters a "real-world-dominated, simulation-constrained" state.
8. The method for predicting the energy efficiency ratio of a central air conditioning chiller unit based on XGBoost and LSTM according to claim 1, characterized in that, The specific method for the incremental learning and model update stage in step S3 is as follows: Step S31, establish a model performance monitoring mechanism: calculate the model's prediction indicators on new real data daily / hourly: MAE, RMSE; The incremental learning and model update phase is configured with a first trigger condition and a second trigger condition. If the first trigger condition and / or the second trigger condition are met, the incremental learning trigger mechanism is initiated. The first trigger condition is defined as the cumulative amount of real data N. real If the proportion exceeds 60%, the second triggering condition is a decrease in model performance. Compare the MAE and RMSE metrics on the validation sets before and after the test. If the RMSE is satisfied... new -RMSE old ≥δ perf If the percentage is ≥10%, the incremental learning mechanism is triggered; where RMSE new The root mean square error (RMSE) of the current model on the latest real-world validation set is used to characterize the model's current COP prediction performance for the latest chiller unit operating conditions. old The root mean square error (RMSE) on the corresponding validation set after the model's last training or incremental learning is used as the baseline value for model performance; δpref is a preset performance degradation threshold. When the increment of RMSE reaches or exceeds this threshold, it is determined that the model performance has degraded, satisfying the second triggering condition for incremental learning; Step S32, Freezing and Unfreezing Configuration of Model Layers: Freeze the XGBoost structural branches; Freeze the LSTM embedding layer; Unfreeze the Attention layer; Unfreeze the gated fusion layer: This allows the fusion weights of "structural features and temporal features" to be adjusted to adapt to feature dependency changes under new data; Unfreeze the output layer: Allows fine-tuning of its parameters to match the COP prediction output pattern of the latest real data; S33, Incremental... Training Data Preparation: Construct a training set from the newly added real data, and simultaneously select independent, up-to-date real data to construct a validation set; S34, Low-Learning-Rate Warm-Start Training: Use "Mean Squared Error of COP Prediction from the Latest Real Data" as the loss function, backpropagate the loss only to the unfrozen layers, and update their parameters; continuously monitor the RMSE of the validation set during training, and terminate fine-tuning when the validation set performance no longer improves after several consecutive rounds; S35, Model Update and Recording: Integrate the fine-tuned unfrozen layer parameters with the frozen layer parameters to obtain the updated hybrid model, which will be used for subsequent chiller unit COP prediction; simultaneously record the current model's RMSE on the validation set as the RMSEold benchmark for the next incremental learning.
9. A computer storage medium storing computer instructions, wherein the computer instructions, when invoked, are used to execute the energy efficiency ratio prediction method for central air conditioning chiller units based on XGBoost and LSTM as described in any one of claims 1-8.
10. An electronic device, wherein, The electronic device includes: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the energy efficiency ratio prediction method for central air conditioning chiller units based on XGBoost and LSTM as described in any one of claims 1-8.