Building load prediction based on hybrid model and start-stop system and method for ground source heat pump

Abstract

This invention discloses a building load forecasting and ground source heat pump start-up and shutdown system and method based on a hybrid model. The method includes: collecting meteorological and hourly load data and preprocessing them; using Pearson, XGBoost, and RF methods for cross-validation feature selection; constructing a VMD-LSTM-RF hybrid forecasting model, reducing data nonstationarity through variational mode decomposition, extracting time-series features using LSTM (long-short time series), and completing nonlinear mapping using RF, and employing a weighted fusion strategy to achieve complementary advantages; executing a graded start-up and shutdown optimization strategy based on the predicted load value, and introducing an equipment balancing protection mechanism. This invention solves the problems of data nonstationarity and nonlinearity in building heating and cooling load forecasting, realizes intelligent start-up and shutdown control of ground source heat pump units, significantly improves forecasting accuracy and equipment operation balance, and reduces system energy consumption and equipment wear.

Description

Technical Field

[0001] This invention relates to the field of building energy conservation and intelligent control technology of HVAC, and in particular to a building load prediction and ground source heat pump start-up and shutdown system and method based on a hybrid model. Background Technology

[0002] Building energy consumption accounts for a significant proportion of my country's total energy consumption, with HVAC systems being a major component, consuming approximately 40%-60% of total building energy. Ground source heat pump systems are widely used in public buildings due to their high efficiency, energy saving, and environmental friendliness. However, the operating efficiency of ground source heat pump systems is closely related to the accurate prediction of building loads. Accurate load prediction is a key prerequisite for realizing intelligent start-up and shutdown of ground source heat pump systems and reducing operating energy consumption.

[0003] Currently, building load forecasting methods mainly fall into two categories: traditional models and machine learning models. Traditional models, such as linear regression (LR) and time series modeling (TSM), are computationally simple but struggle to effectively handle the non-stationarity and nonlinear characteristics of load data. Single machine learning models, such as support vector machines (SVM) and random forests (RF), have limitations in handling complex time-series relationships; while long short-term memory neural networks (LSTM) can capture long-term dependencies in time series data well, their ability to decompose and process non-stationary data is insufficient.

[0004] For unit start-up and shutdown control, existing technologies mostly rely on fixed thresholds or manual experience for scheduling, lacking dynamic optimization strategies that effectively link with load forecasting. Firstly, insufficient load forecasting accuracy leads to frequent unit start-ups and shutdowns or delayed responses. Secondly, start-up and shutdown strategies do not consider the balance of cumulative equipment operating time, resulting in some units operating at high loads for extended periods, leading to accelerated wear and tear, while other units remain idle, severely impacting equipment lifespan and system stability.

[0005] Therefore, there is an urgent need for an integrated technical solution that can simultaneously address the issues of load forecasting accuracy and unit start-up and shutdown optimization, thereby achieving closed-loop coordination of forecasting, decision-making, and control. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned shortcomings by providing a building load forecasting and ground source heat pump start-up and shutdown system and method based on a hybrid model. This addresses the significant non-stationarity and nonlinearity of building heating and cooling load data, which makes it difficult for a single forecasting model to effectively handle both characteristics simultaneously, resulting in insufficient forecasting accuracy. Furthermore, it solves the problem that existing feature selection methods, often based on single indicators, are prone to missing important features or introducing redundant features, affecting model generalization performance. Additionally, it addresses the lack of effective linkage between existing ground source heat pump unit start-up and shutdown strategies and load forecasting, which prevents the realization of multi-unit coordinated operation and equipment balance protection.

[0007] This invention is achieved through the following scheme: A building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model includes a data acquisition module for collecting meteorological environmental parameters and historical building cooling and heating load data, and performing preprocessing. The feature selection module is used to perform cross-validation using Pearson correlation coefficient method, XGBoost feature importance analysis and random forest feature importance analysis to determine the optimal input features for the cold and heat load prediction model; The hybrid prediction module includes a VMD decomposition unit, an LSTM time series feature extraction unit, an RF nonlinear mapping unit, and a weighted fusion unit, which are used to output the predicted building heating and cooling load values ​​for the next time step. The start-stop decision module is used to execute a graded start-stop strategy according to the load forecast value in three load ranges: low, medium, and high, and to prioritize the scheduling of auxiliary units based on the cumulative running time. The balancing protection module is used to monitor the continuous operating time and load rate of the unit, and automatically switch the operating status of similar auxiliary units when the rotation conditions are met.

[0008] Based on the above-mentioned building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model, in the hybrid prediction module, the VMD decomposition unit decomposes the original non-stationary load sequence into K=4 IMF components; the LSTM time series feature extraction unit adopts a two-layer 64-neuron structure with a dropout rate of 0.2; the RF nonlinear mapping unit and the weighted fusion unit adopt a random forest model with n_estimators=700 and max_depth=25.

[0009] Based on the above-mentioned building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model, the start-up and shutdown decision module is configured with ground source heat pump units and magnetic levitation centrifugal chillers on the cooling side, and ground source heat pump units and air-cooled heat pump units on the heating side. Each unit supports five-level variable frequency operation at 0%, 25%, 50%, 75%, and 100%.

[0010] Based on the above-mentioned building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model, the rotation conditions of the balance protection module are: the continuous operating time of the auxiliary unit is >24h and the current load rate is ≥75%, and another auxiliary unit of the same type is in a shutdown state.

[0011] Based on the above-mentioned building load forecasting and ground source heat pump start-up and shutdown system based on a hybrid model, the evaluation index of the forecasting model includes the coefficient of determination R. 2 Root mean square error (RMSE) and mean absolute error (MAE); validation set R for the cooling load prediction model. 2 ≥0.90, validation set R of heat load prediction model2 ≥0.97.

[0012] This solution also discloses a method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model, including the following steps: S1. Data Acquisition and Preprocessing: Collect building meteorological environmental parameters and historical heating and cooling load data; perform linear interpolation of missing values, outlier identification and correction, and Min-Max normalization on the collected data; scale the data to the range of [0,1]. S2. Feature selection through multi-method cross-validation: Pearson correlation coefficient method, XGBoost feature importance analysis and RF feature importance analysis are used to cross-validate the correlation between meteorological parameters and building cooling and heating loads; the input features of the cooling load prediction model and the heating load prediction model are determined by combining the screening results of the three methods. S3. Construct a VMD-LSTM-RF hybrid prediction model: Input the historical load sequence into the Variational Mode Decomposition (VMD) module, set the number of decomposition modes K=4, and obtain multiple Intrinsic Mode Function (IMF) components with specific sparsity characteristics; input each IMF component and the selected meteorological features into the LSTM network for time series feature extraction, and simultaneously input them into the RF model for nonlinear mapping; use a weighted fusion strategy to fuse the outputs of LSTM and RF, and output the predicted building heating and cooling load value Q for the next time step; S4. Unit start-up and shutdown optimization decision; Based on the predicted cooling and heating load value Q, execute a graded start-up and shutdown strategy according to the system operation mode: the cooling mode and the heating mode are respectively divided into three intervals: low load, medium load and high load; in different intervals, start and stop the ground source heat pump unit, magnetic levitation centrifugal chiller unit or air-cooled heat pump unit according to the preset logic, and adjust the frequency conversion gear. S5. Equipment Balance Protection; A unit rotation operation protection mechanism is set up. When any auxiliary unit has been running continuously for more than 24 hours and the current load rate is not lower than 75%, while another auxiliary unit of the same type is in a shutdown state, the system automatically switches the operating unit to realize equipment rotation. The ground source heat pump unit is a basic peak-shaving device and does not participate in the balance rotation.

[0013] The specific method for cross-validation in step S2 is as follows: calculate the Pearson correlation coefficient between each meteorological parameter and the cold and heat load respectively, and select the top three features with the highest |r| threshold; use the XGBoost model to calculate the importance score of each feature based on gain; use the random forest model to calculate the average importance of each feature based on the reduction of impurity; verify the consistency of the screening results of the three methods, and take the intersection feature as the model input.

[0014] The parameter settings for VMD decomposition in step S3 include: number of decomposition modes K=4, penalty factor α=2000.

[0015] The training parameters of the LSTM network in step S3 include: the optimizer is Adam, the initial learning rate is 0.001, the early stopping strategy is adopted, the patience value is 15, and the maximum number of training rounds is 80.

[0016] The weight coefficients of the weighted fusion strategy described in step S3 are determined by testing different weight combinations on the validation set, with the optimization objective being to minimize the root mean square error (RMSE).

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: (1) Significantly improved prediction accuracy: Data stabilization preprocessing is achieved through VMD decomposition, combined with LSTM time series feature extraction and RF nonlinear mapping, and a weighted fusion strategy is adopted to achieve complementary advantages. The cold load prediction model validation set R 2 The R value of the heat load prediction model validation set reached 0.907. 2 The score reached 0.978, which is better than both single models and existing combined models. (2) Feature selection is more reliable: Three different methods are used for cross-validation, which avoids the limitations of a single method, effectively reduces input redundancy, and improves the generalization ability of the model; (3) More intelligent unit scheduling: Based on the hierarchical start-up and shutdown strategy of accurate load forecasting, the coordinated operation of multiple units and accurate load matching are realized, avoiding frequent start-up and shutdown of units and ineffective operation, and effectively reducing system energy consumption; (4) More balanced equipment operation: The introduction of cumulative running time priority scheduling and equipment rotation protection mechanism effectively balances the running time of similar auxiliary units, extends the overall life of equipment, and reduces operation and maintenance costs. Attached Figure Description

[0018] Figure 1 This is a diagram of the VMD-LSTM-RF hybrid prediction model architecture of the present invention; Figure 2 This is a schematic diagram of the VMD-LSTM-RF hybrid prediction model of the present invention; Figure 3 This is a flowchart of the feature engineering screening process for multi-method cross-validation in this invention; Figure 4 This is a flowchart of the staged start-up and shutdown optimization strategy for the cold and heat source unit of the present invention; Figure 5 This invention predicts the cooling / heating load versus the actual cooling / heating supply. Figure 6 The operating conditions of each unit in this invention; Figure 7 The cooling / heating provided by each unit of this invention. Detailed Implementation

[0019] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0020] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0021] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a predetermined orientation, or be constructed and operated in a predetermined orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0023] Example 1 like Figures 1-2 As shown, the present invention provides a technical solution: A building load forecasting and ground source heat pump start-up / shutdown system based on a hybrid model includes: The data acquisition module is used to collect meteorological environmental parameters and historical building heating and cooling load data, and to perform preprocessing. The feature selection module is used to perform cross-validation using Pearson correlation coefficient method, XGBoost feature importance analysis and random forest feature importance analysis to determine the optimal input features for the cold and heat load prediction model; The hybrid prediction module includes a VMD decomposition unit, an LSTM time series feature extraction unit, an RF nonlinear mapping unit, and a weighted fusion unit, which are used to output the predicted building heating and cooling load values ​​for the next time step. The start-stop decision module is used to execute a graded start-stop strategy according to the load forecast value in three load ranges: low, medium, and high, and to prioritize the scheduling of auxiliary units based on the cumulative running time. The balancing protection module is used to monitor the continuous operating time and load rate of the unit, and automatically switch the operating status of similar auxiliary units when the rotation conditions are met.

[0024] In the hybrid prediction module, the VMD decomposition unit decomposes the original non-stationary load sequence into K=4 IMF components; the LSTM temporal feature extraction unit adopts a two-layer 64-neuron structure with a Dropout rate of 0.2; and the RF nonlinear mapping unit and weighted fusion unit adopt a random forest model with n_estimators=700 and max_depth=25.

[0025] In the start / stop decision module, the cooling side is equipped with ground source heat pump units and magnetic levitation centrifugal chillers, and the heating side is equipped with ground source heat pump units and air-cooled heat pump units. Each unit supports five-level variable frequency operation: 0%, 25%, 50%, 75%, and 100%.

[0026] The rotation conditions for the equalization protection module are: the continuous operating time of the auxiliary unit is >24h and the current load rate is ≥75%, and another auxiliary unit of the same type is in a shutdown state.

[0027] The evaluation metrics for the prediction model include the coefficient of determination R. 2 Root mean square error (RMSE) and mean absolute error (MAE); validation set R for the cooling load prediction model. 2 ≥0.90, validation set R of heat load prediction model 2 ≥0.97.

[0028] Example 2 like Figures 1-4 As shown, the present invention provides a technical solution: A method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model includes the following steps: S1. Data Acquisition and Preprocessing: Collect building meteorological environmental parameters and historical heating and cooling load data. The meteorological environmental parameters include dry-bulb temperature, relative humidity, global horizontal radiation, wind speed, wind direction, and cloud cover. Perform linear interpolation for missing values, outlier identification and correction, and Min-Max normalization on the collected data. Scale the data to the range of [0,1]. S2. Feature Selection Through Multi-Method Cross-Validation: Pearson correlation coefficient method, XGBoost feature importance analysis, and RF feature importance analysis were used to cross-validate the correlation between meteorological parameters and building cooling and heating loads. Based on the combined results of these three methods, the input features for the cooling load prediction model were determined to be dry-bulb temperature, global horizontal radiation, and relative humidity (3 items in total); the input features for the heating load prediction model were dry-bulb temperature, global horizontal radiation, wind direction, wind speed, and relative humidity (5 items in total). S3. Construct a VMD-LSTM-RF hybrid prediction model: input the historical load sequence into the Variational Mode Decomposition (VMD) module, set the decomposition mode number K=4, to obtain a plurality of Intrinsic Mode Function (IMF) components with specific sparse characteristics; input each IMF component and the screened meteorological features into the LSTM network respectively for time series feature extraction, wherein the LSTM network adopts a two-layer structure, each layer contains 64 neurons, and the Dropout discard rate is set to 0.2; at the same time, input each IMF component after VMD decomposition and the original meteorological features into the Random Forest (RF) model for nonlinear mapping, wherein the RF model is optimized for parameters via RandomizedSearchCV, with the number of decision trees n_estimators=700 and the maximum tree depth max_depth=25; adopt a weighted fusion strategy to fuse the outputs of LSTM and RF, where the weight of LSTM in the cooling load prediction model is 0.15 and the weight of RF is 0.85, and the weight of LSTM in the heating load prediction model is 0.55 and the weight of RF is 0.45, to output the predicted value of the building cooling and heating load at the next moment; S4. Optimal start-stop decision for units: Based on the predicted cooling and heating load value Q, execute a hierarchical start-stop strategy according to the system operation mode: In the cooling supply mode: when Q ≤ 585kW, only the ground source heat pump unit is started and adjusted with variable frequency according to actual demand; when 585kW < Q ≤ 2,168kW, the ground source heat pump unit operates at 100% load rate, the magnetic suspension centrifugal chiller unit with shorter cumulative operation time is preferentially started, and the variable frequency gear is adjusted according to the ratio [(Q-585) / 1,583]; when Q > 2,168kW, both the ground source heat pump unit and one magnetic suspension centrifugal chiller unit operate at 100% load, and the second magnetic suspension centrifugal chiller unit is put into operation according to the ratio [(Q–585-1,583) / 1,583]; In the heating supply mode: when Q ≤ 585kW, only the ground source heat pump unit operates; when 585kW < Q ≤ 1,127kW, the ground source heat pump unit operates at 100% load, the air-cooled heat pump unit with shorter cumulative operation time is preferentially started, and variable frequency adjustment is performed according to the ratio [(Q-585) / 542]; when Q > 1,127kW, both the ground source heat pump unit and the first air-cooled heat pump unit operate at 100% load rate, and the second air-cooled heat pump unit is put into operation according to the ratio [(Q–585-542) / 542]; S5. Equipment equalization protection: Set a unit rotation operation protection mechanism. When any auxiliary unit has been in continuous operation for more than 24 hours and the current load rate is not lower than 75%, while another auxiliary unit of the same type is in a shutdown state, the system automatically switches the operating units to realize equipment rotation; the ground source heat pump unit, as the basic peak regulation equipment, does not participate in the equalization rotation.

[0029] The specific method for cross-validation in step S2 is as follows: calculate the Pearson correlation coefficient between each meteorological parameter and the cold and heat load respectively, and select the top three features with the highest |r| threshold; use the XGBoost model to calculate the importance score of each feature based on gain; use the random forest model to calculate the average importance of each feature based on the reduction of impurity; verify the consistency of the screening results of the three methods, and take the intersection feature as the model input.

[0030] The parameter settings for VMD decomposition in step S3 include: number of decomposition modes K=4, penalty factor α=2000.

[0031] The training parameters of the LSTM network in step S3 include: the optimizer is Adam, the initial learning rate is 0.001, the early stopping strategy is adopted, the patience value is 15, and the maximum number of training rounds is 80.

[0032] The weight coefficients of the weighted fusion strategy described in step S3 are determined by testing different weight combinations on the validation set, with the optimization objective being to minimize the root mean square error (RMSE).

[0033] The core of this embodiment lies in using a VMD-LSTM-RF hybrid model to address the nonstationarity and nonlinearity issues in building load forecasting, and in optimizing unit start-up and shutdown based on accurate forecasting. Its basic principle is as follows: (1) Variational Mode Decomposition (VMD) Principle: By solving the constrained variational problem, the non-stationary original load sequence is adaptively decomposed into K IMF components with specific center frequencies and finite bandwidths. Each IMF component corresponds to a fluctuation component at a different frequency scale, effectively reducing the non-stationarity and complexity of the data and laying the foundation for subsequent modeling. In this invention, K=4 and the penalty factor α=2000 are set.

[0034] (2) LSTM Temporal Feature Extraction Principle: Long Short-Term Memory (LSTM) networks effectively solve the gradient vanishing problem in traditional recurrent neural networks by introducing three gating mechanisms: forget gate, input gate, and output gate, enabling them to capture long-term dependencies in time series. This invention constructs a two-layer LSTM network, each layer containing 64 neurons, and introduces a Dropout layer (dropout rate 0.2) before the fully connected layer to prevent overfitting. The LSTM network extracts temporal features from each IMF component after VMD decomposition, uncovering deep dynamic patterns in the load data.

[0035] (3) RF Nonlinear Mapping Principle: RF is an ensemble learning algorithm based on the ideas of Bootstrap Aggregating and random feature subspaces. By averaging the prediction results of multiple decision trees, it effectively reduces the variance of a single model. This model has good robustness to noise and can handle nonlinear relationships and interactions between high-dimensional features. The RF model receives the time-series features extracted by LSTM and the original meteorological features, and completes the nonlinear mapping from input features to load values.

[0036] (4) Weighted Fusion Principle: To fully leverage the temporal feature capture capability of LSTM and the stability advantage of RF, this invention employs a weighted fusion strategy to construct a VMD-LSTM-RF hybrid prediction model. Different weight combinations are tested on the validation set, with minimizing RMSE as the optimization objective. Ultimately, the LSTM weight in the cold load prediction model is determined to be 0.15, and the RF weight to be 0.85; while the LSTM weight in the hot load prediction model is 0.55, and the RF weight to be 0.45. This strategy achieves a complementary advantage between temporal feature capture capability and nonlinear mapping capability, resulting in an ensemble learning effect where 1+1>2.

[0037] (5) Staged Start-up and Shutdown Principle: Based on the predicted load value, the system operation is divided into three load ranges: low, medium, and high, according to the cooling / heating mode. In the low load range, only the ground source heat pump unit is activated to meet the basic peak-shaving requirements; in the medium load range, auxiliary units are added and the frequency converter is adjusted proportionally; in the high load range, multiple units operate in coordination to ensure cooling / heating requirements. Through the cumulative operating time priority scheduling mechanism, auxiliary units with shorter cumulative operating times are activated first to achieve accurate load matching and reasonable scheduling of units.

[0038] (6) Balance Protection Principle: To avoid uneven equipment wear caused by long-term continuous operation of a single auxiliary unit, this invention sets up an automatic rotation mechanism based on two indicators: continuous operating time and load rate. When one of the same type of auxiliary units has been running continuously for more than 24 hours with a load rate ≥75%, while the other is in a shutdown state, the system automatically switches the operating unit to achieve equipment rotation. The ground source heat pump unit, as a basic peak-shaving device, does not participate in the balance rotation to ensure the continuity of system power supply.

[0039] Example 3 This embodiment takes a public building in a certain region as an example. The building has a total height of 60m, including 14 floors above ground and 1 floor underground, with a total building area of ​​54,143m². It uses a ground source heat pump system as the heat source and the total air-conditioned area is about 28,388m².

[0040] I. System Equipment Configuration: The summer cooling side is equipped with one ground-source heat pump unit with a rated cooling capacity of 585kW and two magnetic levitation centrifugal chiller units with a rated cooling capacity of 1,583kW. The winter heating side is equipped with one ground-source heat pump unit with a rated heating capacity of 585kW and two air-cooled heat pump units with a rated heating capacity of 542kW. Each unit supports five-level variable frequency operation: 0%, 25%, 50%, 75%, and 100%. The system's maximum cooling capacity is 3,751kW, and its maximum heating capacity is 1,669kW.

[0041] II. Data Acquisition and Preprocessing: The data collected in this embodiment includes two categories: (1) meteorological environment data, which are from the typical meteorological year (2005) data of EnergyPlus, including dry bulb temperature, relative humidity, global horizontal radiation, wind speed, wind direction and overall cloud cover; (2) hourly cooling and heating load data of buildings throughout the year, which are obtained by establishing a fully parameterized calculation model through EnergyPlus building energy consumption simulation software, with a total of 8,760 sets of data throughout the year.

[0042] The preprocessing process includes: filling missing values ​​using linear interpolation; identifying and correcting abnormal data points based on equipment acquisition point anomalies and business logic; and scaling continuous variables to the [0,1] range using Min-Max standardization to eliminate the influence of units and accelerate model convergence. After preprocessing, the datasets are divided into training and test sets in an 8:2 ratio, and sequential slicing is used to ensure the continuity of the time series. The cold / hot load training and test sets contain 7,008 and 1,752 samples, respectively.

[0043] III. Feature selection using multi-method cross-validation: We used three methods—Pearson correlation coefficient method, XGBoost feature importance analysis, and random forest feature importance analysis—to cross-validate the correlation between six candidate meteorological parameters and building heating and cooling loads.

[0044] Pearson analysis results show that the three factors most strongly correlated with cooling load are dry-bulb temperature (r=0.710), global horizontal radiation (r=0.552), and relative humidity (r=-0.269); the three factors most strongly correlated with heat load are dry-bulb temperature (r=0.540), wind speed (r=-0.096), and global horizontal radiation (r=-0.091).

[0045] XGBoost analysis results show that dry-bulb temperature (0.4758) and global horizontal radiation (0.3850) contribute significantly to cooling load, while dry-bulb temperature (0.3993), global horizontal radiation (0.2123), and wind speed (0.1198) are the main influencing factors of heat load.

[0046] Random forest analysis results show that global horizontal radiation (0.283) and dry-bulb temperature (0.273) are of high importance to cooling load, while dry-bulb temperature (0.641) dominates the building heat load.

[0047] Based on the cross-validation results of the three methods, the cooling load prediction model was determined to use three features: dry-bulb temperature, global horizontal radiation, and relative humidity; while the heat load prediction model was determined to use five features: dry-bulb temperature, global horizontal radiation, wind direction, wind speed, and relative humidity.

[0048] IV. Construction of VMD-LSTM-RF Hybrid Prediction Model: The historical load sequence was input into the VMD module for decomposition. The number of decomposition modes K=4 and the penalty factor α=2000 were set to obtain 4 IMF components with different frequency scales.

[0049] The LSTM network employs a two-layer structure, with each layer containing 64 neurons. A Dropout layer is introduced before the fully connected layer, with a dropout rate of 0.2. The Adam optimizer is used, with an initial learning rate of 0.001. An early stopping strategy is employed during training, with a patience value of 15, and a maximum of 80 training epochs. The LSTM network extracts temporal features for each IMF component separately.

[0050] The RF model employs the RandomizedSearchCV strategy to optimize among 150 parameter combinations, ultimately determining the following model parameters: number of decision trees n_estimators = 700, maximum tree depth max_depth = 25, maximum feature ratio max_features = 0.6, and minimum number of split samples per node min_samples_split = 5. The RF model receives time-series features extracted by LSTM and raw meteorological features, completing the nonlinear mapping.

[0051] A weighted fusion strategy was used to construct a VMD-LSTM-RF hybrid prediction model. Different weight combinations were tested on the validation set, with minimizing RMSE as the optimization objective. Finally, the LSTM weight for the cooling load prediction model was determined to be 0.15 and the RF weight to be 0.85, while the LSTM weight for the heating load prediction model was determined to be 0.55 and the RF weight to be 0.45.

[0052] Model evaluation results: For the cooling load prediction model, the R² of the training set is 0.997172, the R² of the validation set is 0.907315, the RMSE is 0.021305, and the MAE is 0.013543; for the heating load prediction model, the R² of the training set is 0.996564, the R² of the validation set is 0.978234, the RMSE is 0.016293, and the MAE is 0.010887. The performance of the two models on the training set and the validation set is close, no significant overfitting occurs, and the generalization ability is good.

[0053] V. Unit start-stop optimization decision-making: The hierarchical start-stop strategy is executed based on the cooling and heating load predicted value Q at the next moment output by the VMD-LSTM-RF model.

[0054] Cooling supply mode: When Q ≤ 585kW, only the ground source heat pump unit is started, and its frequency conversion gear is adjusted according to the actual load demand; when 585kW < Q ≤ 2,168kW, the ground source heat pump unit operates at 100% load rate, the magnetic suspension centrifugal chiller unit with shorter cumulative operation time is preferentially started, and the frequency conversion gear is adjusted according to the ratio of [(Q-585) / 1,583]; when Q > 2,168kW, both the ground source heat pump unit and one magnetic suspension centrifugal chiller unit operate at 100% load, and the second magnetic suspension centrifugal chiller unit is put into operation according to the ratio of [(Q–585-1,583) / 1,583].

[0055] Heating supply mode: When Q ≤ 585kW, only the ground source heat pump unit operates; when 585kW < Q ≤ 1,127kW, the ground source heat pump unit operates at 100% load, the air-cooled heat pump unit with shorter cumulative operation time is preferentially started, and frequency conversion adjustment is performed according to the ratio of [(Q-585) / 542]; when Q > 1,127kW, both the ground source heat pump unit and the first air-cooled heat pump unit operate at 100% load rate, and the second air-cooled heat pump unit is put into operation according to the ratio of [(Q-585-542) / 542].

[0056] VI. Equipment equalization protection: A unit rotation operation protection mechanism is set. When any magnetic suspension centrifugal chiller unit or air-cooled heat pump unit operates continuously for more than 24 hours with the current load rate not lower than 75%, while another unit of the same type is in a shutdown state, the system automatically switches the operating units to realize equipment rotation. As the basic peak regulation equipment of the system, the ground source heat pump unit does not participate in the equalization operation strategy.

[0057] VII. Verification results: To verify the effectiveness of the above strategy, 168 consecutive hours (one week) of cooling and heating load forecast data were selected for strategy verification. The verification results show that the start-stop optimization strategy of this embodiment has good load tracking capability and equipment balance, the deviation of cooling / heating supply on demand is controllable, and the operating time of each device is evenly distributed.

[0058] The correspondences between the predicted cooling / heating load and the actual cooling capacity at each time point are as follows: Figure 5 As shown, the operating conditions of each unit and the timing distribution of cooling / heating supply are as follows: Figure 6 and Figure 7 As shown.

[0059] The verification results show that the start-stop optimization strategy has good load tracking capability and equipment balance, and can provide decision support for the economic operation and equipment management of cold and heat source systems.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A building load prediction and ground source heat pump start-up / shutdown system based on a hybrid model, characterized in that: include: The data acquisition module is used to collect meteorological environmental parameters and historical building heating and cooling load data, and to perform preprocessing. The feature selection module is used to perform cross-validation using Pearson correlation coefficient method, XGBoost feature importance analysis and random forest feature importance analysis to determine the optimal input features for the cold and heat load prediction model; The hybrid prediction module includes a VMD decomposition unit, an LSTM time series feature extraction unit, an RF nonlinear mapping unit, and a weighted fusion unit, which are used to output the predicted building heating and cooling load values ​​for the next time step. The start-stop decision module is used to execute a graded start-stop strategy according to the load forecast value in three load ranges: low, medium, and high, and to prioritize the scheduling of auxiliary units based on the cumulative running time. The balancing protection module is used to monitor the continuous operating time and load rate of the unit, and automatically switch the operating status of similar auxiliary units when the rotation conditions are met.

2. The building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model as described in claim 1, characterized in that: In the hybrid prediction module, the VMD decomposition unit decomposes the original non-stationary load sequence into K=4 IMF components; the LSTM temporal feature extraction unit adopts a two-layer 64-neuron structure with a Dropout rate of 0.2; and the RF nonlinear mapping unit and weighted fusion unit adopt a random forest model with n_estimators=700 and max_depth=25.

3. The building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model as described in claim 1, characterized in that: In the start / stop decision module, the cooling side is equipped with ground source heat pump units and magnetic levitation centrifugal chillers, and the heating side is equipped with ground source heat pump units and air-cooled heat pump units. Each unit supports five-level variable frequency operation: 0%, 25%, 50%, 75%, and 100%.

4. The building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model as described in claim 1, characterized in that: The rotation conditions for the equalization protection module are: the continuous operating time of the auxiliary unit is >24h and the current load rate is ≥75%, and another auxiliary unit of the same type is in a shutdown state.

5. A building load prediction and ground source heat pump start-up and shutdown system based on a hybrid model as described in claim 1, characterized in that: The evaluation metrics for the prediction model include the coefficient of determination R. 2 Root mean square error (RMSE) and mean absolute error (MAE); validation set R for the cooling load prediction model. 2 ≥0.90, validation set R of heat load prediction model 2 ≥0.

97.

6. A method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model, characterized in that: Includes the following steps: S1. Data Acquisition and Preprocessing: Collect building meteorological environmental parameters and historical heating and cooling load data; perform linear interpolation of missing values, outlier identification and correction, and Min-Max normalization on the collected data; scale the data to the range of [0,1]. S2. Feature selection through multi-method cross-validation: Pearson correlation coefficient method, XGBoost feature importance analysis and RF feature importance analysis are used to cross-validate the correlation between meteorological parameters and building cooling and heating loads; the input features of the cooling load prediction model and the heating load prediction model are determined by combining the screening results of the three methods. S3. Construct a VMD-LSTM-RF hybrid prediction model: Input the historical load sequence into the Variational Mode Decomposition (VMD) module, set the number of decomposition modes K=4, and obtain multiple Intrinsic Mode Function (IMF) components with predetermined sparsity characteristics; input each IMF component and the selected meteorological features into the LSTM network for time series feature extraction, and simultaneously input them into the RF model for nonlinear mapping; use a weighted fusion strategy to fuse the outputs of LSTM and RF, and output the predicted building heating and cooling load value Q for the next time step; S4. Unit start-up and shutdown optimization decision; Based on the predicted cooling and heating load value Q, execute a graded start-up and shutdown strategy according to the system operation mode: the cooling mode and the heating mode are respectively divided into three intervals: low load, medium load and high load; in different intervals, start and stop the ground source heat pump unit, magnetic levitation centrifugal chiller unit or air-cooled heat pump unit according to the preset logic, and adjust the frequency conversion gear. S5. Equipment Balance Protection; A unit rotation operation protection mechanism is set up. When any auxiliary unit has been running continuously for more than 24 hours and the current load rate is not lower than 75%, while another auxiliary unit of the same type is in a shutdown state, the system automatically switches the operating unit to realize equipment rotation. The ground source heat pump unit is a basic peak-shaving device and does not participate in the balance rotation.

7. The method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model as described in claim 6, characterized in that: The specific method for cross-validation in step S2 is as follows: calculate the Pearson correlation coefficient between each meteorological parameter and the cold and heat load respectively, and select the top three features with the highest |r| threshold; use the XGBoost model to calculate the importance score of each feature based on gain; use the random forest model to calculate the average importance of each feature based on the reduction of impurity; verify the consistency of the screening results of the three methods, and take the intersection feature as the model input.

8. The method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model as described in claim 7, characterized in that: The parameter settings for VMD decomposition in step S3 include: number of decomposition modes K=4, penalty factor α=2000.

9. The method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model as described in claim 8, characterized in that: The training parameters of the LSTM network in step S3 include: the optimizer is Adam, the initial learning rate is 0.001, the early stopping strategy is adopted, the patience value is 15, and the maximum number of training rounds is 80.

10. The method for building load prediction and ground source heat pump start-up and shutdown based on a hybrid model as described in claim 9, characterized in that: The weight coefficients of the weighted fusion strategy described in step S3 are determined by testing different weight combinations on the validation set, with the optimization objective being to minimize the root mean square error (RMSE).

Images