A multi-mechanism cooling load forecasting method and system

Abstract

The application discloses a multi-mechanism cold supply load prediction method and system, which establishes a machine learning model based on load physical composition and containing a sub-model cooperation prediction system, and a historical data matching model that can construct a hybrid model system with self-defined parameters and adaptive matching algorithms, and automatically selects and runs an adaptive adjustment model for actual prediction. The model self-improvement mechanism is also deployed for the machine learning model and the historical data matching model, respectively, to ensure that the model has self-improvement characteristics during operation, which will ensure that the two models are continuously optimized and always use the better model, significantly improving the scene coverage rate and the ability of the model after data accumulation.

Description

Technical Field

[0001] This invention relates to the fields of building operation and maintenance and artificial intelligence technology, specifically to a multi-mechanism cooling load prediction method and system for large venues. Background Technology

[0002] In recent years, rapid urbanization has spurred the continuous development of the construction industry, leading to its continuous expansion. From 2001 to 2015, total building energy consumption increased significantly. For example, in 2015, the total commercial energy consumption of building operations reached 864 million tons of electricity (tce), accounting for approximately 20% of the country's total energy consumption. Building energy consumption includes lighting, electrical outlets, power, and heating, ventilation, and air conditioning (HVAC) systems, with HVAC systems accounting for the largest share and being one of the key focuses of current building energy conservation efforts.

[0003] Studies have shown that the high energy consumption of HVAC systems is mainly due to the units operating at low load rates, resulting in low operating efficiency of the units and pumps. Operation and maintenance personnel, relying on experience, often adopt high-flow-rate, low-temperature-difference system operation modes, increasing pump energy consumption. Therefore, reasonable operation and adjustment strategies are needed to improve the energy efficiency of HVAC systems. The building's heating and cooling loads are the purpose and basis for the operation and adjustment of the air conditioning system. By studying and analyzing historical building data and establishing predictive models based on this, accurate predictions of future building loads can be made. By determining future cooling needs in advance, reasonable operation strategies for the HVAC system can be arranged to reduce system energy consumption.

[0004] Most existing cooling load forecasting methods are designed for scenarios such as office buildings or commercial complexes. These scenarios have continuous and relatively regular operating conditions, are easy to accumulate data on, and have relatively stable loads. Therefore, it is relatively easy to build machine learning models based on historical data for load forecasting.

[0005] However, in application scenarios such as large venues, there are significant differences in load on weekdays, event days, or other different working conditions. The load is discontinuous in the time dimension and presents an extremely discrete state, which makes data accumulation difficult. When there is a lack of data and the working conditions fluctuate greatly, the implementation of the prediction model based on data training is generally not effective.

[0006] For example, the prediction method mentioned in CN202411117891.9 uses the concept of historical data average and some parameters to make relatively simple predictions, and its prediction objects are building types with continuous working conditions such as commercial, office, and school buildings.

[0007] Therefore, existing methods for predicting cooling loads have the following limitations when applied to large venues:

[0008] 1. Strong dependence on large-scale data, limiting applicability. Existing cooling load forecasting methods are highly dependent on large-scale data. They often require a certain amount of historical data that meets quality requirements for model training and application. However, projects in the early stages or those that have not yet established a data system are unlikely to provide large-scale data, which greatly limits their applicability.

[0009] 2. The model is too simplistic, lacks diversity, and has poor adaptability. Using a single machine learning model or other algorithm for prediction will inevitably result in lower performance for specific tasks under limited conditions, especially for more general models. Furthermore, the adaptability of a single model is significantly limited across different scenarios.

[0010] 3. The model cannot iterate and improve itself. Once the existing prediction model or method is initially established, it is no longer modified or upgraded, or upgrades require a high degree of manual intervention, such as retraining or parameter adjustment, which greatly increases the cost of model iteration and is not conducive to model improvement and efficient iteration.

[0011] 4. Inability to independently select models to maximize their effectiveness. As data accumulates or the scenario changes, the effectiveness of different models will also change. If models are not rotated in a timely manner, a single model may fall into a local optimum.

[0012] Therefore, the present invention aims to solve the aforementioned technical problem by providing a multi-mechanism cooling load prediction method. Summary of the Invention

[0013] To address the aforementioned technical problems with existing cooling load forecasting methods, this invention aims to provide a multi-mechanism cooling load forecasting method. This method constructs a hybrid model system by establishing a machine learning model and a historical data matching model. It also employs automated, optimal operation and adaptive adjustment of model switching, and deploys a model self-improvement mechanism to ensure the model's self-improvement during operation, significantly improving iteration efficiency. Furthermore, it provides a forecasting system capable of implementing this multi-mechanism cooling load forecasting method, effectively overcoming the problems existing in the prior art.

[0014] To achieve the above objectives, the present invention provides a multi-mechanism cooling load forecasting method, which includes the following steps:

[0015] S1: Collect data and build a database in chronological order.

[0016] This step involves collecting raw data on the project and equipment, converting the collected data into a structured entry library, and then storing it in a non-relational database.

[0017] S2: Select the prediction model based on the data scale and characteristics.

[0018] This step verifies the integrity of parameters and the scale of data by combining the physical model with the types and quantities of parameters collected in S1. Based on the integrity of parameters and the scale of data, different prediction models are selected and constructed.

[0019] S3: Predictive Model Building

[0020] This step constructs a machine learning prediction model and a historical data matching model based on the prediction model selected in S2.

[0021] S4: Predictive Model Validation and Debugging

[0022] This step is used to simulate and debug the prediction model built in S3;

[0023] S5: Predictive Model Self-Improvement Mechanism

[0024] This step is used to build a new dataset by adding valid data to the dataset and to verify the prediction performance of the new model compared with the original model, retaining the model with better performance;

[0025] S6: Automated optimal operation of predictive models

[0026] This step is used to compare the root mean square error of the predicted load and the actual load between the two models built in S3 when they are run simultaneously, and to select the best model based on the error.

[0027] Furthermore, if the parameters in S2 are incomplete, it is considered whether there are substitute parameters for the missing parameters. If substitute parameters exist, they can be used to replace the original parameters and added to the subsequent model building process.

[0028] Furthermore, in S3, if the scale of data verified in S2 meets the training requirements of the machine learning prediction model, that is, it can ensure that the model has a certain accuracy after training, then a hybrid system of machine learning prediction model and historical data matching model can be used to predict cooling load; when the parameters are incomplete and there are no substitute parameters or the data scale is small, the historical data matching model can be used.

[0029] Furthermore, the machine learning prediction model includes three sub-models: a total cooling load prediction model, an hourly cooling load prediction model, and an hourly cooling load correction model. All three sub-models are trained using the same training set format, meaning that each sample in the training set contains features and labels. The labels for the total cooling load prediction model and the hourly cooling load correction model are the total cooling load and the cooling load correction value for the next hour, both of which are single values. The label for the hourly cooling load prediction model is the hourly cooling load, which is in list form and contains the cooling load for all hours of the day, totaling 24 data points.

[0030] Furthermore, the construction of a historical data matching model includes two parts: the basic model architecture and the matching algorithm;

[0031] Basic model architecture: First, a candidate parameter set needs to be established. Select at least two parameters from the candidate parameter set and assign them weight coefficients between 0 and 1. The sum of the weight coefficients assigned to all parameters should be equal to 1. After completing the parameter selection and weight coefficient assignment, solidify this configuration to form the basic model architecture.

[0032] Matching algorithm: Based on the corresponding parameter input selected in the basic model architecture, the parameter information to be predicted for the day is used to execute the matching algorithm.

[0033] Furthermore, the matching algorithm consists of 8 steps:

[0034] M2-S1: Retrieves structured data from the database constructed in S1;

[0035] M2-S2: Traverse the data: Calculate the absolute difference between all data entries and the corresponding numerical parameters in the parameter information for the predicted day;

[0036] M2-S3: Based on the absolute difference calculated for each parameter, each data entry is sorted in descending order and then normalized and scored.

[0037] M2-S4: When the parameters of the basic model architecture include a date type, this parameter will be scored when the date type of the predicted day matches the date type of the data entry.

[0038] M2-S5: In M2-S3 and M2-S4, the scores for relevant parameters, including date type, have been calculated. Based on the score information and the weight coefficient of each parameter in the basic model architecture, the total score is calculated for each data entry.

[0039] M2-S6: When the parameters of the basic model architecture include a date, select the top 10 data items with the highest total score in M2-S5, calculate the similarity between the date and the predicted date, and sort them according to similarity. The score of the most similar item is used, and so on until the last item.

[0040] M2-S7: Based on the scores of the 10 data entries, calculate the total score for each data entry according to the weight coefficient of each parameter in the basic model architecture; the total score here will cover the total score in M2-S5;

[0041] M2-S8: The cooling load information from the data entry with the highest total score is selected as the predicted value, and the remaining data entries will be used for the self-improvement mechanism.

[0042] Furthermore, M2-S4 and M2-S6 in the above steps are optional steps and can only be executed if the basic model architecture contains date types and date parameters. If M2-S6 is not executed, M2-S7 also does not need to be executed. The remaining steps are executed in sequence, and steps that do not need to be executed can be skipped.

[0043] To achieve the above objectives, the present invention provides a multi-mechanism cooling load prediction system, comprising:

[0044] A database creation module is used to collect actual operation and maintenance information and equipment data of specific projects or engineering projects and build an original database.

[0045] A prediction model selection module interacts with the database construction module and is used to select a prediction model based on the data scale and characteristics in the database construction module.

[0046] A prediction model building module, which interacts with the prediction model selection module, is used to build machine learning prediction models and historical data matching models.

[0047] The prediction model verification and debugging module interacts with the prediction model construction module to test and debug the constructed prediction model.

[0048] The prediction model self-improvement mechanism module interacts with the prediction model verification and debugging module. It constructs a new dataset by adding valid data to the dataset and verifies the prediction effect of the new model and the original model.

[0049] The predictive model automated optimization operation module interacts with the predictive model construction module. It is used to compare the root mean square error of the predicted load and the actual load of the two models when the two models are running simultaneously and to perform optimization operation.

[0050] The multi-mechanism cooling load prediction method provided by this invention has the following features and advantages:

[0051] The hybrid model system no longer relies excessively on data scale and supports load forecasting under low-volume data. It constructs a hybrid model system by establishing a machine learning model and a historical data matching model. The machine learning model is suitable for scenarios with a certain scale of data that is relatively complete and of high quality, while the historical data matching model can adapt to situations such as the early stages of project operation and maintenance when the amount of data accumulated is small.

[0052] A multi-model system with different capabilities and adaptability was constructed. A machine learning model based on load physics was established, which includes a prediction system with sub-models, and a historical data matching model with customizable parameters and adaptive matching algorithms.

[0053] A self-improvement mechanism was established for the model. Self-improvement mechanisms were developed for both the machine learning model and the historical data matching model, allowing them to continuously improve themselves during operation to enhance actual prediction performance and eliminating the need for manual intervention in subsequent processes.

[0054] An automated optimization system for the hybrid model has been established. During operation, the two models are automatically compared in real time based on criteria, and the model with better performance is switched for actual prediction. Coupled with a self-improvement mechanism, this ensures that the two models are continuously optimized and that the better-performing model is always used, significantly improving scenario coverage and the adaptability of the model after data accumulation. Attached Figure Description

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] Figure 1 This is a schematic diagram of the overall process of the multi-mechanism cooling load forecasting method.

[0057] Figure 2 This is a schematic diagram of the database construction process in this multi-mechanism cooling load forecasting method;

[0058] Figure 3 This is a schematic diagram illustrating the selection process of the prediction model in this multi-mechanism cooling load prediction method;

[0059] Figure 4 This is a flowchart illustrating the execution of M1 in this multi-mechanism cooling load forecasting method.

[0060] Figure 5 This is a schematic diagram of the process for constructing M2 in the multi-mechanism cooling load forecasting method.

[0061] Figure 6 This is a flowchart illustrating the M1 self-improvement mechanism in the multi-mechanism cooling load forecasting method.

[0062] Figure 7 This is a flowchart illustrating the M2 self-improvement mechanism in the multi-mechanism cooling load forecasting method.

[0063] Figure 8 This is a flowchart illustrating the automatic optimal operation process in the multi-mechanism cooling load forecasting method.

[0064] Figure 9 This is a schematic diagram of the M1 training implementation case in the multi-mechanism cooling load forecasting method.

[0065] Figure 10 This is a schematic diagram of the M2 training implementation case in the multi-mechanism cooling load forecasting method. Detailed Implementation

[0066] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0067] To address the aforementioned technical problems with existing cooling load forecasting methods, the present invention aims to provide a multi-mechanism cooling load forecasting method. This method constructs a hybrid model system by establishing a machine learning model and a historical data matching model, and automatically optimizes and adaptively adjusts model switching through optimal operation. It also deploys a model self-improvement mechanism to ensure that the model has self-improvement characteristics during operation, thereby greatly improving iteration efficiency.

[0068] The multi-mechanism cooling load prediction method provided by this invention is described in [reference needed]. Figure 1 It includes the following steps:

[0069] S1: Collect data and build a database in chronological order (e.g., ... Figure 2 )

[0070] This step involves collecting raw data about the project and equipment and converting the collected data into a structured entry library to build a database for easy retrieval and retrieval later.

[0071] S11: Collect and organize data to form raw data

[0072] This step is used to collect two parts of data: actual operation and maintenance information and equipment data for specific projects or engineering projects.

[0073] The operation and maintenance information includes, but is not limited to, the date, the activities or date type (such as weekday, event day, weekend, etc.), the number of people, indoor and outdoor temperature, humidity, etc.

[0074] Equipment data includes, but is not limited to, data related to HVAC systems, such as inlet and outlet water temperatures and ice storage capacity. These data generally originate from sensors, recording devices, or integrated building automation (BA) systems within the HVAC system and are primarily used to calculate total cooling load and hourly cooling load.

[0075] S12: Convert raw data into structured entries

[0076] This step is used to convert the raw data collected in S11 into structured data entries (in the form of "label: value"), where the label is the individual data name in S11 (such as "date") and the value is the specific data value (such as "2024-09-20").

[0077] S13: Structured data entries are stored in a non-relational database.

[0078] This step is used to store the transformed structured entries in a non-relational database (such as MongoDB) to maintain the designed data structure and facilitate subsequent retrieval and retrieval.

[0079] S2: Select the prediction model based on the data scale and characteristics (e.g., ...). Figure 3 )

[0080] S21: Extract the parameter types and quantity of data collected in S1.

[0081] S22: Perform parameter integrity verification in conjunction with the physical model. The composition of the load physical model is shown below:

[0082] Q τ =KF1(t τ-ξ -t n )+F1X g J wτ +1000(G1+G2)(H τ -H n )+n(q s +q l )+W τ

[0083] Where K is the heat transfer coefficient of the building envelope, W / m 2 / K; F1 is the area of ​​the enclosure structure, in m² 2 ;t τ-ξ The temperature for calculating the cooling load at the moment when the temperature wave is applied is ℃; t n F2 represents the indoor temperature (°C); F2 represents the area of ​​the exterior windows and glass curtain wall (m²). 2 ;X g J is the construction correction factor for exterior windows and glass curtain walls; wτ To calculate the cooling load intensity of solar radiation through the glass without shading at the specified time, in W / m². 2 G1 represents the organized fresh air volume, kg / s; G2 represents the unorganized infiltration air volume, kg / s; H τ To calculate the outdoor air enthalpy at the given time, kJ / kg; H n The value of indoor air enthalpy at the given time is given in kJ / kg; n is the number of people in the room; qs For sensible heat generation of personnel, W; q l Latent heat of personnel, W; W τ The electrical energy consumed by all equipment within a building, representing the heat generated by the equipment.

[0084] S23: Validation of Data Scale

[0085] Based on the above S22, parameter integrity is verified in conjunction with the physical model. If the parameters are complete, the next step of data scale verification can be performed. If the parameters are incomplete, it is considered whether there are substitute parameters for the missing parameters. If substitute parameters exist, they can be used to replace the original parameters and added to the subsequent model building process. When the parameters are complete or substitute parameters exist, data scale verification is performed (S23).

[0086] When the data scale meets the training requirements of the machine learning prediction model, that is, it can ensure that the model has a certain accuracy after training (which can be judged according to the specific situation, usually the scale needs to be higher than hundreds), the machine learning prediction model M1 and the historical data matching model M2 can be used to form a hybrid system for cooling load prediction. When the parameters are incomplete and there are no substitute parameters or the data scale is small, the historical data matching model M2 can be used.

[0087] S3: Predictive Model Building

[0088] This step is used to build the prediction model. There are two main types of prediction models: machine learning prediction model M1 and historical data matching model M2. M1 and M2 are built using different methods and need to be built separately.

[0089] Among them, such as Figure 4 The machine learning prediction model M1 contains three sub-models: the total cooling load prediction model M11, the hourly cooling load prediction model M12, and the hourly cooling load correction model M13.

[0090] All three seed models are trained using the same training set format, where each sample in the training set contains features and labels. M11 and M13 are labeled with the total cooling load and the cooling load compliance correction value for the next hour, both of which are single values. M12 is labeled with the hourly cooling load, which is in list form and contains the cooling load for all hours of the day, totaling 24 data points.

[0091] M11 and M13 can be trained using methods such as linear regression, nonlinear regression, neural networks, XGBoost, LightGBM, and random forests, while M12, due to its requirement for an output list, can only be trained using methods such as neural networks, XGBoost, LightGBM, and random forests.

[0092] The training method involves retrieving the required data from the database, dividing it into training and test sets according to a 9:1, 8:2, or other ratio, and then selecting an algorithm to train models M11, M12, and M13 respectively. After training, the models are tested on the test set until the accuracy requirements are met. Metrics used to evaluate model performance or accuracy include cross-entropy loss and mean squared error. Their expressions are as follows:

[0093] Cross-entropy loss:

[0094]

[0095] Mean square error:

[0096]

[0097] Once the sub-models of M1 have been trained, they can be used for actual load forecasting through a fixed process.

[0098] First, the input needs to be based on the characteristics required by M11 and M12, i.e., the characteristics used during training. M11 will output the total cooling load, while M12 will output the hourly cooling load. The total cooling load is compared with the sum of the hourly cooling loads. When the sum of the hourly loads is greater than the total load, the hourly load is used; when the sum of the hourly loads is less than the total load, the hourly load is calibrated with the total load, that is, the total load is redistributed according to the hourly load ratio under the original hourly load to determine the final hourly load. In the actual operation of the HVAC system, if there is operating data from the previous hour, the cooling load for the next hour is corrected through M13, and its input characteristics are consistent with the training set characteristics input during training.

[0099] like Figure 5 The construction of the historical data matching model M2 includes two parts: the basic model architecture and the matching algorithm.

[0100] First: The establishment of the basic model architecture requires the establishment of a candidate parameter set, which includes, but is not limited to, parameters such as date, date type, outdoor maximum temperature, outdoor maximum humidity, outdoor average temperature, and indoor average temperature.

[0101] The above parameters are divided into two categories: special parameters and numeric parameters. Special parameters include dates and date types, while the remaining parameters are numeric parameters.

[0102] Then, select at least two parameters from the candidate parameter set and assign them weight coefficients between 0 and 1. The sum of the weight coefficients assigned to all parameters should equal 1. After completing the parameter selection and weight coefficient assignment, solidify this configuration to form the basic model architecture.

[0103] Based on the corresponding parameters selected in the basic model architecture, the parameter information for the day to be predicted is input, and a matching algorithm is executed based on this parameter information. The matching algorithm consists of 8 steps, namely M2-S1 to M2-S8:

[0104] M2-S1: To obtain structured data from the database built in S1, the program needs to call all valid data from the database. If the accumulated data has been accumulated for more than 3 years, then the data from the last 3 years can be retrieved.

[0105] M2-S2: Iterates through the data and calculates the absolute difference between all data entries and the corresponding numerical parameters in the parameter information for the predicted day. For example, the basic model architecture includes the parameter "average outdoor temperature". The database has a data entry for "2024-09-19" with an average outdoor temperature of 30.2. The predicted average outdoor temperature for the current day is 30.5, so the absolute difference is 0.3. This process is repeated to calculate the absolute difference between all parameters in all entries and the parameters for the predicted day.

[0106] M2-S3: Based on the absolute difference calculated for each parameter, each data entry is sorted in descending order and then normalized and scored. The scores for each sorting are as follows:

[0107]

[0108] Where si is the score of the data entry corresponding to the i-th parameter, ni is the ranking of the data entry among all data entries according to the absolute difference of the i-th parameter, and L is the number of data entries.

[0109] M2-S4: When the parameters of the basic model architecture include a date type, the score for this parameter is 10 points when the date type of the predicted day is consistent with the date type of the data entry; if a custom similarity rule is used, such as when the date type is a rehearsal day and a general event, they are considered similar, and the score for this parameter is 5 points; otherwise, the score for the date type of the data entry is 0 points.

[0110] M2-S5: In M2-S3 and M2-S4, the scores for relevant parameters, including date type, have already been calculated. Based on the score information and the weight coefficient of each parameter in the basic model architecture, the total score is calculated for each data entry.

[0111] M2-S6: When the parameters of the basic model architecture include a date, the top 10 data items with the highest total score in M2-S5 are selected, the similarity between the date and the predicted date is calculated, and the items are sorted according to similarity. The most similar item scores 10 points, and so on until the last item scores 1 point. The similarity is calculated by prioritizing the distance between the month and day, and then considering the distance between the year.

[0112] M2-S7: Based on the scores of the 10 data entries, calculate the total score for each data entry according to the weight coefficient of each parameter in the basic model architecture; the total score here will cover the total score in M2-S5.

[0113] M2-S8: The cooling load information from the data entry with the highest total score is selected as the predicted value, and the remaining data entries will be used for the self-improvement mechanism.

[0114] In the above steps, M2-S4 and M2-S6 are optional steps. They can only be executed if the basic model architecture contains date types and date parameters. If M2-S6 is not executed, M2-S7 also does not need to be executed. The remaining steps are executed in sequence, and steps that do not need to be executed can be skipped.

[0115] S4: Predictive Model Validation and Debugging

[0116] This step is used to test and debug the prediction model built in S3.

[0117] Once the model in S3 is built, it needs to be put into actual operation for testing. The first 72 hours are used for simulation and prediction, and the root mean square error between the actual hourly cooling load and the predicted hourly cooling load is calculated by comparing the actual cooling load value. A threshold is set, and the model is considered to have passed the verification and debugging only when it is less than the threshold.

[0118] S5: Predictive Model Self-Improvement Mechanism

[0119] This step is used for the self-improvement of the prediction model and the enhancement of its predictive performance. The machine learning prediction model M1 and the historical data matching model M2 follow different self-improvement mechanisms.

[0120] like Figure 6 The core of the self-improvement mechanism of the machine learning prediction model M1 is to add new effective data to the dataset as data is gradually accumulated, train a new model with the newly constructed dataset, verify the prediction effect of the new model and the original model, and retain the model with better performance.

[0121] First, a training set increment threshold Δm needs to be set. Second, the amount of newly added valid data Δn in the database needs to be checked periodically. When Δn ≥ Δm, the self-improvement mechanism of M1 is triggered. New samples are added to the original training set to form a new training set, and the new model is trained according to the original mechanism to obtain a new M1 (including 3 sub-models).

[0122] Compare the performance of the new model and the original model on the same test set. If the new model is worse than the original model, you can choose to discard the data Δn, or add an additional data review mechanism to further improve the data quality, and return to the step of periodically checking the database for newly added valid data Δn, and continue the loop.

[0123] When the new model is superior to the original model, the new model is added to the workflow and runs in "shadow mode" for 72 hours. "Shadow mode" means that the new model also makes predictions of cooling load at the same time, but its prediction data does not participate in the actual control of the equipment.

[0124] After 72 hours, compare the average deviations of the new model and the original model from the actual cooling load. If the new model is still better than the original model, then the new model will replace the original model. However, the original model needs to be backed up and saved in case of unforeseen circumstances requiring version rollback. If the new model is worse than the original model, then it will be handled in the same way as the test set performance of the original model.

[0125] like Figure 7 The core of the historical data matching model M2 self-improvement mechanism is to adjust the weight ratio between various parameters to ensure that the data entries most similar to the actual cooling load can obtain the highest possible total score.

[0126] First, the M2 model runs for 72 hours, with each cycle lasting 24 hours. In each cycle, the root mean square error (MSE) of the hourly cooling load and the actual hourly cooling load is calculated for the 10 data entries (including the data entry with the highest total score) retained in M2-S8. The average value of the MSE for each cycle is taken as the final comparison index. The data entry with the smallest MSE value for the cycle is selected. If the data entry with the smallest MSE value for the cycle is the original data entry with the highest total score, no changes are made, and the cycle continues for the next 72 hours.

[0127] When the data entry with the smallest mean MSE in a given period is another data entry, compare the scores of each parameter of the data entry with the highest total score with the data entry with the smallest mean MSE in the given period. Increase the weight coefficient of the parameter with the lower score in the data entry with the smallest mean MSE in the given period, for example, by 0.1 or 0.05. After that, normalize the weights of all parameters to ensure that their sum equals 1. This completes the weight coefficient update of the basic model architecture, and the next round of iteration can be executed.

[0128] S6: Automated optimal execution of predictive models (e.g.) Figure 8 )

[0129] Considering that the accuracy of machine learning prediction model M1 and historical data matching model M2 may fluctuate under different scenarios and data scales, a mechanism is needed to automatically select the better model in order to ensure that the model with better prediction performance can be used in a timely manner.

[0130] When both models are built, they are run simultaneously. Using 72 hours of running data as a benchmark, the root mean square error (MSE) of the predicted load and the actual load of models M1 and M2 is compared. Subsequently, the model with the smaller MSE is switched as the actual load prediction model, while the other model runs synchronously and continuously improves itself. This process is repeated to complete the optimal selection. When only M2 is available, no optimization is required.

[0131] The multi-mechanism cooling load forecasting method presented in this example can be configured into a corresponding software program to form a multi-mechanism cooling load forecasting system. When running, this software program will execute the aforementioned multi-mechanism cooling load forecasting method and store it in a suitable storage medium for the processor to retrieve and execute.

[0132] The resulting multi-mechanism cooling load forecasting system mainly includes the following functional modules: database establishment module, forecasting model selection module, forecasting model construction module, forecasting model verification and debugging module, forecasting model self-improvement mechanism module, and forecasting model automated optimal operation module.

[0133] The database creation module in this system is used to collect data and build the database in chronological order. The database module includes a collection unit, a transformation unit, and a storage unit.

[0134] The collection unit is used to collect actual operation and maintenance information and equipment data for specific projects or engineering projects and organize them to form raw data.

[0135] The transformation unit interacts with the collection unit to convert the raw data collected by the collection unit into structured entries.

[0136] The storage unit interacts with the transformation unit to store the structured data entries transformed by the transformation unit into a non-relational database.

[0137] The database creation module is configured to execute the above database creation steps to achieve the corresponding functions.

[0138] The prediction model selection module in this system interacts with the database construction module. It is used to select a prediction model based on the data scale and characteristics. The prediction model selection module includes an extraction unit, a verification unit, and a data scale verification unit.

[0139] The extraction unit is used to extract the types and quantities of parameters collected in the database creation module.

[0140] The verification unit interacts with the extraction unit, and performs parameter integrity verification in conjunction with the physical model based on the type and quantity of the extracted parameters.

[0141] The data scale verification unit interacts with the verification unit to verify whether the relevant parameters in the physical model verification data are complete.

[0142] The prediction model selection module is configured to perform the above prediction model selection steps to achieve the corresponding functions.

[0143] The prediction model building module and the prediction model selection module in this system interact with each other to build machine learning prediction models and historical data matching models.

[0144] The prediction model building module is configured to execute the above prediction model building steps to achieve the corresponding functions.

[0145] The prediction model verification and debugging module in this system interacts with the prediction model building module to test and debug the constructed prediction model.

[0146] The prediction model verification and debugging module is configured to perform the above prediction model verification and debugging steps to achieve the corresponding functions.

[0147] The prediction model self-improvement mechanism module and the prediction model verification and debugging module in this system interact with each other. The latter constructs a new dataset by adding valid data to the dataset and verifies the prediction effect of the new model and the original model, retaining the model with better performance.

[0148] The prediction model self-improvement mechanism module is configured to execute the above prediction model self-improvement mechanism steps to achieve the corresponding functions.

[0149] The prediction model automatic optimization operation module in this system interacts with the prediction model construction module. It is used to compare the root mean square error (MSE) of the predicted load and the actual load when the two models are running at the same time, and then switch to the model with the smaller MSE as the actual load prediction model.

[0150] The automated optimization module for the prediction model is configured to execute the aforementioned automated optimization steps to achieve the corresponding functions.

[0151] The multi-mechanism cooling load forecasting method composed of the above schemes is illustrated below with examples to show its working process in specific applications. It should be noted that the working process here is only for illustrative purposes and does not constitute a limitation on this scheme.

[0152] The implementation scheme of this invention is described using the training of machine learning prediction model M1 and the construction of historical data matching model M2 as the object.

[0153] like Figure 9M1 includes three sub-models: M11, M12, and M13. During training, training sets and test sets with different features and labels need to be constructed according to the original data.

[0154] One implementation scheme is that the training and test sets of M11 include weighted parameters with features such as date type, average outdoor temperature, maximum outdoor temperature, average outdoor humidity, maximum outdoor humidity, minimum outdoor humidity, average temperature of n indoor measuring points, number of people indoors, number of fresh air units turned on and their opening degree, and labeled as total cooling load.

[0155] The training and test sets of M12 include features such as date type, outdoor hourly temperature, outdoor hourly humidity, outdoor main wind direction and hourly wind speed, indoor hourly temperature at n measuring points, indoor hourly number of people, and indoor hourly air intake. The parameters with hourly data are all in list format, containing data for each hour. The label is the hourly cooling load, which is also in list format.

[0156] The training and testing sets of M13 include the following features: actual cooling load of the previous hour, average outdoor temperature of the previous hour, average outdoor humidity of the previous hour, average wind speed of the main outdoor wind direction of the previous hour, average temperature of n indoor measuring points of the previous hour, average number of people indoors of the previous hour, predicted cooling load of the next hour (optional), average outdoor temperature of the next hour, average outdoor humidity of the next hour, average wind speed of the main outdoor wind direction of the next hour, and average number of people indoors of the next hour. The label is the cooling load correction value of the next hour.

[0157] The three models have different inputs and outputs and different dimensions. Algorithms such as neural networks can be used to build models with consistent usage and properties.

[0158] like Figure 10 One implementation scheme for building the M2 model is to establish a set of candidate parameters including date, date type, average outdoor temperature, maximum outdoor temperature, average outdoor humidity, maximum outdoor humidity, and average indoor temperature. From this set, five parameters are selected: date, date type, average outdoor temperature, average indoor temperature, and maximum outdoor temperature. These parameters are then assigned weight coefficients of 0.10, 0.40, 0.15, 0.15, and 0.20, respectively, to build the basic model architecture.

[0159] After inputting the information for the predicted day, the matching algorithm is executed on the data stored in the database to obtain the final matching result.

[0160] In summary, the multi-mechanism cooling load forecasting method and system provided in this solution no longer overly rely on data scale. It has adaptable solutions for different data scales and is also suitable for low-scale load forecasting, providing a comprehensive load forecasting solution for the entire project operation and maintenance cycle.

[0161] Simultaneously, a machine learning model based on load physics and incorporating sub-models and a prediction system was established, along with a historical data matching model that allows for customizable parameters and adaptive matching algorithms. It exhibits significant universality, requiring only minor adjustments to adapt to a wide range of applications.

[0162] Secondly, a self-improvement mechanism for the model has been established, which can automatically optimize the model's prediction performance without human intervention after deployment.

[0163] In addition, an automatic optimization system for hybrid models was established. Different models are applicable to different scenarios, ensuring that the better-performing model can be used for actual prediction in real time.

[0164] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-mechanism cooling load forecasting method, characterized in that, The multi-mechanism cooling load forecasting method includes the following steps: S1: Collect data and build a database in chronological order. This step involves collecting raw data on the project and equipment, converting the collected data into a structured entry library, and then storing it in a non-relational database. S2: Select the prediction model based on the data scale and characteristics. This step verifies the integrity of parameters and the scale of data by combining the physical model with the types and quantities of parameters collected in S1. Based on the integrity of parameters and the scale of data, different prediction models are selected and constructed. If the parameters are complete, perform data scale verification; if the parameters are incomplete, consider whether there are substitute parameters for the missing parameters. If substitute parameters exist, they can be used to replace the original parameters and added to the subsequent model building process; when the parameters are complete or substitute parameters exist, perform data scale verification. If the data scale meets the training requirements of the machine learning prediction model, a hybrid system of machine learning prediction model and historical data matching model can be used to predict cooling load. If the parameters are incomplete and there are no substitute parameters or the data scale is small, the historical data matching model should be used. S3: Predictive Model Building This step constructs a machine learning prediction model and a historical data matching model based on the prediction model selected in S2. S4: Predictive Model Validation and Debugging This step is used to simulate and debug the prediction model built in S3; S5: Predictive Model Self-Improvement Mechanism This step is used to build a new dataset by adding valid data to the dataset and to verify the prediction performance of the new model compared with the original model, retaining the model with better performance; S6: Automated optimal operation of predictive models This step is used to compare the root mean square error of the predicted load and the actual load between the two models built in S3 when they are run simultaneously, and to select the best model based on the error.

2. The multi-mechanism cooling load forecasting method according to claim 1, characterized in that, The machine learning prediction model includes three sub-models: total cooling load prediction model, hourly cooling load prediction model, and hourly cooling load correction model. All three sub-models are trained using the same training set format, meaning that each sample in the training set contains features and labels. The labels for the total cooling load prediction model and the hourly cooling load correction model are the total cooling load and the correction value for the next hour's cooling load, both of which are single values. The labels for the hourly cooling load prediction model are the hourly cooling load, which is in list form and contains the cooling load for all hours of the day, totaling 24 data points.

3. The multi-mechanism cooling load prediction method according to claim 1, characterized in that, The construction of a historical data matching model consists of two parts: the basic model architecture and the matching algorithm. Basic model architecture: First, a candidate parameter set needs to be established. Select at least two parameters from the candidate parameter set and assign them weight coefficients between 0 and 1. The sum of the weight coefficients assigned to all parameters should be equal to 1. After completing the parameter selection and weight coefficient assignment, solidify this configuration to form the basic model architecture. Matching algorithm: Based on the corresponding parameter input selected in the basic model architecture, the parameter information to be predicted for the day is used to execute the matching algorithm.

4. The multi-mechanism cooling load prediction method according to claim 3, characterized in that... ； The matching algorithm consists of 8 steps: M2-S1: Retrieves structured data from the database constructed in S1; M2-S2: Traverse the data: Calculate the absolute difference between all data entries and the corresponding numerical parameters in the parameter information for the predicted day; M2-S3: Based on the absolute difference calculated for each parameter, each data entry is sorted in descending order and then normalized and scored. M2-S4: When the parameters of the basic model architecture include a date type, this parameter will be scored when the date type of the predicted day matches the date type of the data entry. M2-S5: In M2-S3 and M2-S4, the scores for relevant parameters, including date type, have been calculated. Based on the score information and the weight coefficient of each parameter in the basic model architecture, the total score is calculated for each data entry. M2-S6: When the parameters of the basic model architecture include a date, select the top 10 data items with the highest total score in M2-S5, calculate the similarity between the date and the predicted date, and sort them according to similarity. The score of the most similar item is used, and so on until the last item. M2-S7: Based on the scores of the 10 data entries, calculate the total score for each data entry according to the weight coefficient of each parameter in the basic model architecture; the total score here will cover the total score in M2-S5; M2-S8: The cooling load information from the data entry with the highest total score is selected as the predicted value, and the remaining data entries will be used for the self-improvement mechanism.

5. A multi-mechanism cooling load forecasting system, characterized in that, It includes: A database creation module is used to collect actual operation and maintenance information and equipment data of specific projects or engineering projects and build an original database. A prediction model selection module interacts with the database construction module and is used to select a prediction model based on the data scale and characteristics in the database construction module. A prediction model building module, which interacts with the prediction model selection module, is used to build machine learning prediction models and historical data matching models. The prediction model verification and debugging module interacts with the prediction model construction module to test and debug the constructed prediction model. The prediction model self-improvement mechanism module interacts with the prediction model verification and debugging module. It constructs a new dataset by adding valid data to the dataset and verifies the prediction effect of the new model and the original model. The predictive model automated optimization operation module interacts with the predictive model construction module. It is used to compare the root mean square error of the predicted load and the actual load of the two models when the two models are running simultaneously and to perform optimization operation.

Images