Thermal load rapid collaborative prediction method based on multi-model dynamic weight fusion

Abstract

The invention discloses a thermal load rapid collaborative prediction method and system based on multi-model dynamic weight fusion, and the method comprises the steps: obtaining the related data of centralized heating based on a data source, building a database, supplementing missing data through employing a Lagrange interpolation algorithm, carrying out the restoration processing of the supplemented data, and obtaining a prediction result; converting the original data into numerical values in the same interval; dividing the original data into a training set and a test set; constructing a multi-model collaborative prediction framework, and designing a dynamic weight fusion mechanism; executing a multi-model collaborative prediction process on the multi-model collaborative prediction framework, carrying out weighted fusion on output results of all models according to dynamic weights, outputting a prediction curve and a confidence interval through a visual platform, constructing a heat supply scheduling optimization scheme based on the prediction results, and combining an economic cost model and a carbon emission constraint condition to obtain a heat supply scheduling optimization scheme. And an optimal operation strategy considering energy efficiency and environmental protection is generated. The method can adapt to the characteristics of nonlinearity, time-varying characteristics and complexity of a thermal load system, and the accuracy and robustness of prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of centralized heating technology, and is applicable to multi-timescale heat load prediction and intelligent scheduling decision-making for centralized heating systems. In particular, it relates to a rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion. Background Technology

[0002] As an important component of the energy system, centralized heating systems account for a significant proportion of energy consumption and carbon emissions. Accurately predicting the heat load of centralized heating systems is crucial for optimizing system operation, improving energy efficiency, and reducing carbon emissions.

[0003] Traditional methods for forecasting centralized heating loads have several limitations. For example, some methods require high data quality; in actual heating processes, equipment failures, sensor malfunctions, and communication interruptions can lead to data anomalies, resulting in significant prediction deviations if these data are used directly for forecasting. Some forecasting models are insufficient in handling complex multivariate relationships and cannot fully consider the various factors affecting heat load, such as outdoor temperature, supply water temperature, return water temperature, and climate factors. Existing methods often struggle to balance the accuracy of short-term and long-term heat load forecasts, failing to meet the planning needs of heating companies at different time scales. Existing single models cannot balance short-term and long-term forecast accuracy: statistical models (such as ARIMA) are effective for short-term forecasts of stationary sequences, but long-term forecasts are susceptible to cumulative errors; grey models (GM) are advantageous for predicting long-term trends in small samples, but lack sensitivity to short-term fluctuations. Furthermore, traditional methods do not effectively combine the complementary characteristics of the two types of models, resulting in limited forecast efficiency and adaptability. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention discloses a rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion, specifically including:

[0005] A database was established based on data sources to obtain relevant data on centralized heating, including real-time temperature information and historical heat load data of each heat exchange station.

[0006] Read the original data from the database, use the Lagrange interpolation algorithm to fill in the missing data, and perform repair processing on the supplemented data to convert the original data into values ​​within the same range;

[0007] The original data is divided into a training set and a test set;

[0008] A multi-model collaborative prediction framework is constructed, which includes multiple prediction models, and a dynamic weight fusion mechanism is designed.

[0009] The multi-model collaborative prediction framework is subjected to a multi-model collaborative prediction process. The output results of each model are dynamically weighted and fused to generate an integrated prediction value covering the short, medium and long term. The prediction curve and confidence interval are output through a visualization platform. Based on the prediction results, a heating scheduling optimization scheme is constructed. Combined with the economic cost model and carbon emission constraints, an optimal operation strategy that takes into account both energy efficiency and environmental protection is generated.

[0010] Furthermore, the multi-model collaborative prediction framework includes:

[0011] The first prediction model is an autoregressive integral moving average model based on time series analysis. The autoregression order p and the moving average order q are determined based on the autocorrelation coefficient ACF and the partial autocorrelation coefficient PACF, and the parameters are optimized using the AIC criterion.

[0012] The second prediction model is a multivariate grey prediction model based on grey system theory. It generates the heat load sequence by first-order accumulation, constructs grey differential equations, and solves the parameters by least squares method to complete the long-term trend prediction.

[0013] The third model compatibility interface is reserved for future expansion interfaces, supporting the integration of neural networks and support vector machine prediction models.

[0014] Furthermore, the dynamic weight fusion mechanism is as follows:

[0015] The weighting strategy based on time scale is as follows: short-term forecasts are given priority and high weights to the first forecasting model, while the weights of the second forecasting model are gradually increased for medium- and long-term forecasts.

[0016] Based on adaptive adjustment of model performance, using mean squared error as the evaluation metric, the weight ratio of each model is dynamically adjusted. The calculation formula is as follows:

[0017]

[0018] Among them, w i For the weights of the i-th model, MSE i This represents the prediction error.

[0019] Furthermore, when establishing the database: firstly, a data priority classification mechanism is established, marking real-time temperature and pressure data of heat exchange stations as primary data, and meteorological elements and historical heat load data as secondary data; multi-source data are synchronized in real time through edge computing nodes, and timestamp alignment method is used to solve the problem of inconsistent time sequence of heterogeneous data.

[0020] Furthermore, to address the issue of abnormal raw data caused by equipment failures and sensor malfunctions during actual heating processes, the Laida method is used for abnormal data detection and removal. When repairing the supplemented data, Lagrange interpolation is used for filling, and normalization or Z-Score methods are used to standardize the training and test sets.

[0021] Furthermore, when constructing the autoregressive integral moving average model: firstly, the range of values ​​for the autoregressive order and the moving average order q are initially determined based on the partial autocorrelation coefficient PAC and the autocorrelation coefficient p. Then, the minimum AIC value is searched using AIC to optimize the sum of values, thus obtaining candidate models for the autoregressive integral moving average model. The candidate models are fitted, and the accuracy of the models is evaluated based on the mean square error criterion. Finally, the model with the smallest MSE value is selected as the final autoregressive integral moving average model.

[0022] Furthermore, when constructing the multivariate grey prediction model: firstly, a feasibility test is performed on the heat load time series, and initial non-negative data series that meet the conditions are selected. Generate a sequence by performing first-order accumulation. Then generate the sequence Z using the adjacent means. 1 ),in

[0023] Construct a data matrix and data vector, and estimate the parameter sequence u = of the grey differential equation using the least squares method.

[0024] ab] T = (B T B) -1 B T Y

[0025] This leads to the prediction model.

[0026] By loading pre-trained model weights through a third-model-compatible interface, a plug-and-play process for new models can be achieved.

[0027] Furthermore, when constructing a dynamic weight fusion mechanism: a smooth transition interval for weights is set, and during the transition from the short-term to the medium-term stage, the model weights are gradually adjusted using linear interpolation to avoid sudden changes in prediction results; a sliding time window mechanism is introduced to periodically update the MSE values ​​of each model to ensure the timeliness of weight allocation.

[0028] A rapid collaborative prediction system for heat load based on multi-model dynamic weight fusion includes:

[0029] Data acquisition module: Used to acquire centralized heating-related data from multiple data sources, including real-time temperature and pressure data of each heat exchange station, outdoor meteorological data, and historical heat load data, and to prioritize the acquisition according to the importance and reliability of the data for heat load prediction.

[0030] The data preprocessing module is used to remove outliers from the collected raw data using the Laida method, supplement missing data using the Lagrange interpolation algorithm, and standardize the data using the Min-max method and logarithmic transformation method in the normalization method to convert the sample data into values ​​within the same range to meet the data specification requirements for subsequent use.

[0031] Model building module: Builds ARIMA model and multivariate grey prediction model, and optimizes parameters and evaluates accuracy for each model;

[0032] Data prediction module: Used to train the constructed model with a large amount of historical data, use the trained model to predict future heat load, provide short-term, medium-term and long-term heat load prediction results, evaluate the results, and determine the applicability of the model in heat load prediction at different time scales.

[0033] By employing the aforementioned technical solutions, this invention provides a rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion. This method prioritizes the ARIMA model for capturing detailed fluctuations in short-term predictions, while gradually increasing the weight of the grey model for medium- and long-term predictions to fit trend changes. Simultaneously, it reserves a model interface to support algorithm expansion. A dynamic weight fusion mechanism is designed, specifically including a time-scale-based weight allocation strategy and adaptive adjustment based on model performance. Multi-model collaborative prediction is executed, fusing the outputs of each model according to dynamic weights to generate an integrated prediction value covering short-, medium-, and long-term periods. The prediction curve and confidence interval are output through a visualization platform. Based on the prediction results, a heating scheduling optimization scheme is constructed. Combining an economic cost model and carbon emission constraints, an optimal operating strategy that balances energy efficiency and environmental protection is generated. Therefore, this method can adapt to the nonlinear, time-varying, and complex characteristics of heat load systems, improving the accuracy and robustness of predictions. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 Flowchart of the method

[0036] Figure 2 Data processing flowchart

[0037] Figure 3 Flowchart for building the Autoregressive Integral Moving Average (ARIMA) model

[0038] Figure 4 Flowchart for constructing a multivariate grey prediction model

[0039] Figure 5 A structural diagram of a centralized heating data processing and model building device Detailed Implementation

[0040] To make the technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention:

[0041] like Figure 1 The method for rapid collaborative prediction of heat load based on multi-model dynamic weight fusion, as shown, specifically includes the following steps:

[0042] S1: Obtain centralized heating-related data from multiple data sources, including but not limited to real-time temperature (primary supply water temperature, primary return water temperature, secondary supply water temperature, secondary return water temperature), pressure data, outdoor meteorological data (temperature, humidity, wind speed, etc.) and historical heat load data of each heat exchange station.

[0043] S2: The raw data read from the database in S1 is processed using the Laida method to remove outlier data, and then the missing data is supplemented using the Lagrange interpolation algorithm. After the data is repaired, the Min-max method and logarithmic transformation method in the normalization method are used for standardization to convert the sample data into values ​​in the same range to meet the data specification requirements for subsequent use.

[0044] S3: Split the dataset into training and test sets.

[0045] S4: Construct a multi-model collaborative prediction framework, including:

[0046] The first prediction model is the Autoregressive Integral Moving Average (ARIMA) model based on time series analysis. The autoregressive order (p) and moving average order (q) are determined based on the autocorrelation coefficient (ACF) and partial autocorrelation coefficient (PACF), and the parameters are optimized using the AIC criterion.

[0047] The second prediction model is a multivariate grey prediction model (GM(1,n)) based on grey system theory. It generates the heat load sequence by first-order accumulation (1-AGO), constructs grey differential equations, solves the parameters by least squares method, and completes the long-term trend prediction.

[0048] Third model compatibility interface: Reserved extension interface to support the integration of other prediction models such as neural networks and support vector machines.

[0049] S5: Design a dynamic weight fusion mechanism, specifically including:

[0050] The weighting strategy based on time scale is as follows: for short-term forecasts (1-7 days), the first forecasting model is given a high weight, while for medium- and long-term forecasts (1-3 months to 1-5 years), the weight of the second forecasting model is gradually increased.

[0051] Adaptive adjustment based on model performance: Using mean squared error (MSE) as the evaluation metric, the weight ratios of each model are dynamically adjusted. The calculation formula is as follows:

[0052]

[0053] Among them, w i For the weights of the i-th model, MSE i Its prediction error.

[0054] S6: Perform multi-model collaborative prediction, dynamically weighting and fusing the outputs of each model to generate an integrated prediction covering short-term, medium-term, and long-term data. The prediction curve and confidence interval are then output through a visualization platform. Based on the prediction results, a heating scheduling optimization scheme is constructed. Combining the economic cost model and carbon emission constraints, an optimal operating strategy that balances energy efficiency and environmental protection is generated.

[0055] Steps S1 / S2 / S3 / S4 / S5 / S6 are executed sequentially;

[0056] Furthermore, the process of obtaining sample data is as follows:

[0057] S11: Establish a data priority classification mechanism, marking real-time temperature and pressure data of heat exchange stations as primary data (collection frequency ≤ 2 minutes), and meteorological elements and historical heat load data as secondary data (collection frequency ≤ 1 hour).

[0058] S12: Real-time synchronization of multi-source data is achieved through edge computing nodes, and timestamp alignment technology is used to solve the problem of inconsistent timing of heterogeneous data.

[0059] Further data processing is performed on the sample data. The data processing flowchart is as follows: Figure 2 As shown, the process is as follows:

[0060] S21: Preprocess the data. For abnormal raw data caused by equipment failure, sensor malfunction, etc. during the actual heating process, use the Raida method (3δ criterion) to detect and remove abnormal data.

[0061] S22: For missing data in S21 that has been processed for abnormal data or that was already present, use Lagrange interpolation to fill in the missing data.

[0062] S23: Considering that the input variables of the heat load prediction model have different dimensions and physical meanings, in order to avoid model failure due to data dimensionality issues, the normalization method or Z-Score method is used to standardize the sample dataset.

[0063] Further: Split the dataset into training and test sets.

[0064] Furthermore: By combining deep learning algorithms, a prediction model is constructed. The flowchart for constructing the ARIMA model is as follows. Figure 3 As shown, the gray model construction flowchart is as follows: Figure 4 The process is as follows:

[0065] S41: Construct an ARIMA model. First, determine the possible range of values ​​for the autoregression order and the moving average order q based on the partial autocorrelation coefficient (PAC) and the autocorrelation coefficient p.

[0066] S42: Then, the minimum AIC value is searched using the Akaike information criterion (AIC), and the sum is further optimized to obtain candidate models such as ARIMA(1,0,0), ARIMA(2,0,0), and ARIMA(7,0,5).

[0067] S43: Fit the candidate models in S42, evaluate the model accuracy according to the mean squared error (MSE) criterion, and select the model with the smallest MSE value as the final ARIMA model.

[0068] S44: Construct a multivariate grey prediction model GM(1,n). First, conduct a feasibility test on the heat load time series to ensure the feasibility of the model.

[0069] S45: For the initial non-negative data sequence that meets the conditions Generate a sequence by performing first-order accumulation. Then generate the sequence Z using the adjacent means. 1 ),in

[0070] S46: Construct the data matrix and data vector, and estimate the parameter sequence u = [ab] of the grey differential equation using the least squares method. T = (B T B) -1 B T Y, thus obtaining the prediction model

[0071] S47: Load pre-trained model weights through a third model compatibility interface to enable plug-and-play functionality for new models.

[0072] Furthermore: A dynamic weight fusion mechanism is designed, prioritizing high weights for the first prediction model in short-term forecasts (1-7 days), and gradually increasing the weights of the second prediction model in medium- to long-term forecasts (1-3 months to 1-5 years). Specifically, this includes:

[0073] S51: Set a weight smoothing transition interval. During the transition from short-term to medium-term (days 7-15), use linear interpolation to gradually adjust the model weights to avoid sudden changes in the prediction results.

[0074] S52: Introducing a sliding time window mechanism to periodically update the MSE values ​​of each model, ensuring the timeliness of weight allocation.

[0075] Furthermore: Multi-model collaborative prediction is performed, and the outputs of each model are dynamically weighted and fused to generate an integrated prediction covering short-term, medium-term, and long-term data. The prediction curve and confidence interval are then output through a visualization platform. Based on the prediction results, a heating scheduling optimization scheme is constructed. Combining the economic cost model and carbon emission constraints, an optimal operating strategy that balances energy efficiency and environmental protection is generated. The specific operations are as follows:

[0076] S61: Input the data from S3 into the pre-built ARIMA model and GM(1,n) model to predict the heat load.

[0077] S62: Using mean square error (MSE) as an evaluation index, calculate the MSE values ​​of the prediction results of the ARIMA model and the GM(1,n) model, and comprehensively evaluate and determine the applicability of the two models in heat load prediction at different time scales.

[0078] S63: Based on the forecast results and the actual operating conditions of the heating system, such as equipment capacity and pipeline transport capacity, an optimized heating dispatching scheme is generated. By establishing economic cost models and environmental impact models, different dispatching schemes are evaluated, and the scheme with the lowest cost and least environmental impact is selected, providing a scientific basis for the decision-making of heating companies.

[0079] like Figure 5 The system shown is a rapid collaborative prediction system for heat load based on multi-model dynamic weight fusion, comprising:

[0080] Data acquisition module: Used to acquire centralized heating-related data from multiple data sources, including real-time temperature (primary supply water temperature, primary return water temperature, secondary supply water temperature, secondary return water temperature), pressure data, outdoor meteorological data (temperature, humidity, wind speed, etc.) and historical heat load data of each heat exchange station, and prioritize the acquisition according to the importance and reliability of the data for heat load prediction.

[0081] The data preprocessing module is used to remove outliers from the collected raw data using the Laida method, then to fill in missing data using the Lagrange interpolation algorithm, and to standardize the data using the Min-max method and logarithmic transformation method in the normalization method, converting the sample data into values ​​within the same range to meet the data specifications required for subsequent use.

[0082] Model building module: Constructs ARIMA model and multivariate grey prediction model GM(1,n), and optimizes the parameters and evaluates the accuracy of each model.

[0083] Data prediction module: Used to train the constructed model with a large amount of historical data, use the trained model to predict future heat load, provide short-term (1-7 days), medium-term (1-3 months) and long-term (1-5 years) heat load prediction results, and evaluate the results to determine the applicability of the model in heat load prediction at different time scales.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion, characterized in that... include: A database was established based on data sources to obtain relevant data on centralized heating, including real-time temperature information and historical heat load data of each heat exchange station. Read the original data from the database, use the Lagrange interpolation algorithm to fill in the missing data, and perform repair processing on the supplemented data to convert the original data into values ​​within the same range; The original data is divided into a training set and a test set; A multi-model collaborative prediction framework is constructed, which includes multiple prediction models, and a dynamic weight fusion mechanism is designed. The multi-model collaborative prediction framework is subjected to a multi-model collaborative prediction process. The output results of each model are dynamically weighted and fused to generate an integrated prediction value covering the short, medium and long term. The prediction curve and confidence interval are output through a visualization platform. Based on the prediction results, a heating scheduling optimization scheme is constructed. Combined with the economic cost model and carbon emission constraints, an optimal operation strategy that takes into account both energy efficiency and environmental protection is generated.

2. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 1, characterized in that: The multi-model collaborative prediction framework includes: The first prediction model is an autoregressive integral moving average model based on time series analysis. The autoregression order p and the moving average order q are determined based on the autocorrelation coefficient ACF and the partial autocorrelation coefficient PACF, and the parameters are optimized using the AIC criterion. The second prediction model is a multivariate grey prediction model based on grey system theory. It generates the heat load sequence by first-order accumulation, constructs grey differential equations, and solves the parameters by least squares method to complete the long-term trend prediction. The third model compatibility interface is reserved for future expansion interfaces, supporting the integration of neural networks and support vector machine prediction models.

3. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 1, characterized in that: The dynamic weight fusion mechanism is as follows: The weighting strategy based on time scale is as follows: short-term forecasts are given priority and high weights to the first forecasting model, while the weights of the second forecasting model are gradually increased for medium- and long-term forecasts. Based on adaptive adjustment of model performance, using mean squared error as the evaluation metric, the weight ratio of each model is dynamically adjusted. The calculation formula is as follows: Among them, w i For the weights of the i-th model, MSE i This represents the prediction error.

4. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 1, characterized in that: When establishing the database: First, a data priority classification mechanism is established, marking real-time temperature and pressure data of heat exchange stations as primary data, and meteorological elements and historical heat load data as secondary data; multi-source data are synchronized in real time through edge computing nodes, and timestamp alignment method is used to solve the problem of inconsistent time sequence of heterogeneous data.

5. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 1, characterized in that: To address the issue of abnormal raw data caused by equipment failures and sensor malfunctions during actual heating processes, the Laida method is used for abnormal data detection and removal. When repairing the supplemented data, Lagrange interpolation is used for filling, and normalization or Z-Score methods are used to standardize the training and test sets.

6. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 2, characterized in that: When constructing an autoregressive integral moving average model: First, determine the range of values ​​for the autoregressive order and the moving average order q based on the partial autocorrelation coefficient PAC and the autocorrelation coefficient p. Then, use AIC to search for the minimum AIC value to optimize the sum of values, thereby obtaining candidate models for the autoregressive integral moving average model. Fit the candidate models, evaluate the model accuracy based on the mean square error criterion, and select the model with the minimum MSE value as the final autoregressive integral moving average model.

7. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 2, characterized in that: When constructing a multivariate grey prediction model: First, a feasibility test is performed on the heat load time series, and initial non-negative data series that meet the conditions are selected. Generate a sequence by performing first-order accumulation. Then generate the sequence Z using the adjacent means. 1 ),in Construct a data matrix and data vector, and estimate the parameter sequence u = of the grey differential equation using the least squares method. a b] T =(B T B) -1 B T Y This leads to the prediction model. By loading pre-trained model weights through a third-model-compatible interface, a plug-and-play process for new models can be achieved.

8. The rapid collaborative prediction method for heat load based on multi-model dynamic weight fusion according to claim 2, characterized in that: When constructing a dynamic weight fusion mechanism: a smooth transition interval for weights is set, and linear interpolation is used to gradually adjust the model weights during the transition from the short-term to the medium-term phase to avoid sudden changes in prediction results; a sliding time window mechanism is introduced to periodically update the MSE values ​​of each model to ensure the timeliness of weight allocation.

9. A rapid collaborative prediction system for heat load based on multi-model dynamic weight fusion, comprising: Data acquisition module: Used to acquire centralized heating-related data from multiple data sources, including real-time temperature and pressure data of each heat exchange station, outdoor meteorological data, and historical heat load data, and to prioritize the acquisition according to the importance and reliability of the data for heat load prediction. The data preprocessing module is used to remove outliers from the collected raw data using the Laida method, supplement missing data using the Lagrange interpolation algorithm, and standardize the data using the Min-max method and logarithmic transformation method in the normalization method to convert the sample data into values ​​within the same range to meet the data specification requirements for subsequent use. Model building module: Builds ARIMA model and multivariate grey prediction model, and optimizes parameters and evaluates accuracy for each model; Data prediction module: Used to train the constructed model with a large amount of historical data, use the trained model to predict future heat load, provide short-term, medium-term and long-term heat load prediction results, evaluate the results, and determine the applicability of the model in heat load prediction at different time scales.

Images