Error size judgment and correction method for heat supply system load prediction point

Abstract

The application discloses a method for judging and correcting error size of a heat supply system load prediction point, and comprises the following steps: S1, obtaining average outdoor air temperature of each day in the future; S2, performing difference on the average outdoor air temperature of the day and the average outdoor air temperature of the previous day, and calculating variation of the average outdoor air temperature of each day in the future; S3, sliding a time window on the difference result of step S2, and dynamically judging a prediction point with large error; S4, introducing an error correction feature epsilon on the basis of an original prediction feature set FS, and obtaining a new prediction feature set FS'; and S5, taking recursive multiple linear regression as a prediction model, and taking the prediction feature set FS' as a prediction feature set, and performing online dynamic prediction on future heat load. Results show that the method can identify a point with large error in advance, and can effectively reduce the error of a local prediction point after introducing the correction feature.

Description

Technical fields:

[0001] This invention relates to the field of heating system load prediction, specifically a method for judging and correcting the error magnitude of heating system load prediction points. Background technology:

[0002] Heat load is an important basis for the operation and regulation of heating systems. Timely and accurate prediction of heat load is of great significance for maintaining the comfort of users' indoor environment, the economical and stable operation of the system, and the achievement of dual carbon targets.

[0003] Current research on heat load forecasting error correction mainly involves making a preliminary forecast of the heat load, studying the historical distribution patterns of forecast errors, and then correcting the preliminary forecast to obtain the final forecast. Current research does not address the reasons for large forecast points and then address the errors at these key points. Furthermore, current evaluation indicators used in heating system load forecasting often focus on the overall forecasting process, rarely paying attention to a small number of particularly large local forecast points. However, these few points with large errors are crucial for the safe and stable operation of the heating system. Therefore, there is an urgent need to study the accuracy improvement of large error points in heat load forecasting.

[0004] Therefore, a method for judging and correcting the error magnitude of the load prediction point of a heating system is invented to solve the above problems. Summary of the Invention:

[0005] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method for judging and correcting the error magnitude of the load prediction point of a heating system.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for judging and correcting the error magnitude of load prediction points in a heating system, the method comprising the following steps:

[0008] To calculate the future daily average outdoor air temperature, obtain the average outdoor air temperature for each day, and then subtract the average outdoor air temperature of the previous day from the average outdoor air temperature of the current day. Obtain the difference results representing the changes in the daily average outdoor air temperature over the next few days;

[0009] Set the threshold for daily average outdoor air temperature change. A time window with a width of p days is set, and this time window slides across the difference results; each time window contains p... If the condition satisfying equation (3) occurs within the time window, the time window is determined to be a key point time window, and the predicted point within the time window is the key point that has been identified.

[0010]

[0011] The original prediction feature set FS for heat load prediction is obtained by using feature selection. An error correction feature ε is introduced on the original prediction feature set FS to obtain a new prediction feature set FS′.

[0012] FS = [X1, X2, ..., X k (4)

[0013] FS′=[X1,X2,…,X k ,ε] (5)

[0014]

[0015] Where FS is the original prediction feature set; X is a single feature; k is the number of features; and ε is the error correction feature. Let q be the predicted heat load for the i-th period, in kW; i-1 The actual heat load value for the (i-1)th cycle, in kW;

[0016] During the dynamic sliding judgment process, when the time window is determined to be a key point time window, the error correction characteristic is the predicted heat load value of the i-th cycle. Compared with the actual heat load q in the (i-1)th cycle i-1 The relative error is determined; otherwise, the error correction feature is set to 0.

[0017] Using recursive multiple linear regression as the prediction model and FS′ as the prediction feature set, we can perform online dynamic prediction of future heat load.

[0018] The features in the original prediction feature set are: the historical heat load of the previous three cycles, the outdoor temperature of the prediction cycle, and the secondary temperature difference of the previous cycle, i.e.

[0019] FS = [q i-1 ,q i-2 ,q i-3 ,t out,i ,Δt 2,i-1 ]

[0020] Where, q i -1,q i-2 ,q i-3 For the historical heat load of the previous three cycles, t out,i To predict the outdoor temperature over a period of time, Δt 2,i-1 This is the second temperature difference of the previous cycle.

[0021] Furthermore, the predicted heat load value for the i-th period Obtained from the prediction model.

[0022] Furthermore, the average daily outdoor air temperature is obtained from formula (1).

[0023]

[0024] Among them, t out,n Let be the outdoor air temperature in the nth hour, in °C; The average daily outdoor air temperature is given in °C.

[0025] Future changes in average daily outdoor air temperature Obtained from formula (2),

[0026]

[0027] in, Let be the average outdoor air temperature on day j, in °C; The average outdoor air temperature on day j-1, in °C;

[0028] Future outdoor air temperature is obtained from weather forecasts.

[0029] Furthermore, k = 3-5; The temperature range is 3℃-5℃; the cycle is the control cycle of the heating station, which is 5-8 hours.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] This invention dynamically identifies key points online by analyzing the daily variation of outdoor air temperature and using a sliding time window. Then, at these key points, the variation amplitude ε of the heat load is introduced as a correction feature into the prediction feature set. Finally, a recursive multiple linear regression prediction model is used to dynamically predict the heat load online, ultimately establishing a method for judging and correcting the error magnitude of heating system load prediction points. This method can successfully identify points with large errors in advance, preparing for targeted error correction later. The prediction error is effectively reduced after introducing the correction feature ε, the predicted heat load is closer to the actual heat load, the number of prediction points with large errors is reduced, and the prediction results for non-key points are not significantly disturbed. Attached image description:

[0032] Figure 1 This is a roadmap of a method for judging and correcting the error magnitude of a heating system load prediction point, as described in this invention.

[0033] Figure 2 This is a schematic diagram of the recursive multiple linear regression prediction model involved in this invention.

[0034] Figure 3 This is an ARE distribution map of all predicted points in this embodiment of the invention;

[0035] Figure 4 This is the difference result of the daily average outdoor air temperature in the embodiments of the present invention;

[0036] Figure 5 This is an ARE distribution map of the identified key points in the embodiments of the present invention;

[0037] Figure 6 This is a comparison diagram before and after error correction in an embodiment of the present invention. Detailed implementation method:

[0038] The present invention will be further explained below with reference to the embodiments and accompanying drawings, but this is not intended to limit the scope of protection of this application.

[0039] The core innovation of this application is to use a sliding time window to identify local time points where large errors will occur in future heat load forecasting by using the difference results of outdoor temperature. Then, an error correction feature is introduced only at these points during forecasting to reduce forecast errors and improve forecast accuracy.

[0040] This invention provides a method for judging and correcting the error magnitude of load prediction points in a heating system (see [link]). Figure 1 This involves calculating the average outdoor air temperature for each day, subtracting the previous day's average outdoor air temperature from the current day's average outdoor air temperature, using a sliding time window to identify key points. Specifically, based on the daily variation in outdoor air temperature, a threshold value is used to dynamically identify prediction points with large errors (referred to as "key points"). Then, error correction features are introduced at these identified prediction points to correct the large errors. These error correction features are then used to dynamically predict future heat loads online. The specific steps are as follows:

[0041] Step S1. Calculate the average daily outdoor air temperature in the future.

[0042]

[0043] Among them, t out,n Let be the outdoor air temperature in the nth hour, in °C.

[0044] Step S2. Calculate the change in the average outdoor air temperature of the next day by subtracting the average outdoor air temperature of the previous day from the average outdoor air temperature of the current day. Outdoor air temperature is the temperature that needs to be predicted in the future, obtained from weather forecasts.

[0045]

[0046] in, Let be the average outdoor air temperature on day j, in °C; The average outdoor air temperature on day j-1 (the day before day j), in °C.

[0047] Step S3. Due to the thermal inertia of large-scale centralized heating systems and buildings, a time window with a width of p days is set (the larger the thermal inertia, the larger p is) and a threshold for the daily average outdoor air temperature change is set. Let the time window slide over the difference results in step S2. Each time window contains p... If the condition satisfying equation (3) occurs within a time window, the time window is determined to be a key point time window, and the predicted point within the time window is the identified key point.

[0048]

[0049] in, The threshold for the daily average outdoor air temperature change is set in advance, in °C.

[0050] Because heating systems and buildings possess thermal inertia and heat storage characteristics, the heat load of heating systems under sudden weather changes differs from the periodic, stable heat load patterns of outdoor weather. Furthermore, the magnitude and duration of sudden changes in outdoor air temperature both affect the pattern of the heating system's heat load. This change in the heat load pattern caused by sudden outdoor weather changes often leads to inaccurate heat load predictions. Therefore, the fundamental reason for the critical point is sudden weather changes.

[0051] When the heating system is running stably, the following heat balance equation exists:

[0052]

[0053] Among them, c p V is the specific heat capacity of air at constant pressure, kJ / (kg·℃); V is the volume of indoor air, m³. 3 ρ is the indoor air density, kg / m³ 3 ;t in τ is the indoor air temperature, °C; q1(τ) is the heat supplied to the building by the heating network, kW; q2(τ) is the heat loss of the building envelope, kW; q3(τ) is the heat gained by disturbance within the building, kW; q4(τ) is the heat transferred by solar radiation, kW; q5(τ) is the heat stored in furniture, interior building envelope, etc., kW; τ is a certain moment.

[0054] The formula for calculating the heat loss q2(τ) of the building envelope is shown in (8):

[0055] q2(τ)=K b F b [t in (τ)-t out(τ)] (8)

[0056] Among them, K b The building's overall heat transfer coefficient, kW / (m²). 2 ·℃); F b The total heat dissipation area of ​​the building is in meters. 2 ;t out (τ) represents the outdoor air temperature, in °C.

[0057] Substituting formula (8) into formula (7) yields:

[0058]

[0059] While ensuring stable indoor comfort, The above formula then becomes:

[0060] q1(τ)=K b F b [t in (τ)-t out (τ)]-q3(τ)-q4(τ)-q5(τ) (10)

[0061] If we ignore the heat loss along the pipeline, the heat load q of the heating station is approximately equal to the heat q1(τ) supplied to the building by the heating pipeline. Then, the heat load q of the heating station can be expressed in the functional form shown in equation (11) as follows:

[0062] q(τ)=f[K b F b , t in (τ), t out (τ), q3(τ), q4(τ), q5(τ)] (11)

[0063] It is evident that the heat load of a heating station is related not only to outdoor air temperature but also to building envelope characteristics, room disturbances, solar radiation, and room heat storage. When outdoor air temperature fluctuates significantly within a short period, the heat load of the heating station differs from the periodic, stable heat load pattern of outdoor weather. This variation in heat load pattern often leads to inaccurate heat load predictions. Therefore, this invention uses the daily variation range of outdoor air temperature to dynamically identify prediction points with large errors online through a variation range threshold. These prediction points are localized points of sudden changes in outdoor temperature, i.e., critical points.

[0064] Step S4. Based on the original predicted feature set FS, introduce the error correction feature ε to obtain a new predicted feature set FS′.

[0065] FS = [X1, X2, ..., X k (4)

[0066] FS′=[X1,X2,…,X k ,ε] (5)

[0067]

[0068] Where FS is the original prediction feature set; X is a single feature; k is the number of features; and ε is the error correction feature. The predicted heat load value for the i-th cycle (obtained by identifying the control cycle, 8h) is kW; q i-1 Let be the actual heat load value for the (i-1)th cycle, in kW. During the dynamic sliding judgment process, if the time window is determined to be a critical point time window, a correction coefficient is introduced; otherwise, the characteristic value is set to 0.

[0069] Step S5. Using recursive multiple linear regression (recursive MLR) as the prediction model and FS′ as the prediction feature set, perform online dynamic prediction of future heat load.

[0070] The matrix expression and expansion of multiple linear regression are as follows:

[0071] Y = X × β + μ (12)

[0072]

[0073] Among them, y i x is the dependent variable; ij There are multiple independent variables; β i μ is the regression coefficient. i This is random error.

[0074] Traditional multiple linear regression models predict the future by using a certain proportion of data as a training set to fit a multiple linear model, and then substituting the independent variables corresponding to the dependent variable to be predicted into the model to predict the unknown dependent variable. This presents a problem: as time goes on, new data is continuously generated, and the previously fitted model may not be applicable to later data. Therefore, this invention incorporates the concept of recursive least squares into the multiple linear regression model.

[0075] Recursive least squares involves feeding data in batches and dynamically updating the model based on new data. Each additional period of data updates the multiple linear regression model for online dynamic prediction. For example... Figure 2As shown, the original data from the first two periods are fitted into the first multiple linear regression model. The input parameters of the third period are then substituted into this model to obtain the predicted value for the third period. The original data from the first three periods are then fitted into the second multiple linear regression model. The input parameters of the fourth period are then substituted into this model to obtain the predicted value for the fourth period. This process is repeated, and the multiple linear regression model is updated for online dynamic prediction with each additional period.

[0076] In this embodiment, the original feature set FS is determined based on feature selection as the historical heat load of the previous three cycles, the outdoor temperature of the predicted cycle, and the secondary temperature difference of the previous cycle:

[0077] FS = [q i-1 ,q i-2 ,q i-3 ,t out,i ,Δt 2,i-1 ]

[0078] Where, q i-1 ,q i-2 ,q i-3 For the historical heat load of the previous three cycles, t out,i , for the outdoor temperature of the predicted period, Δt 2,i-1 This is the second temperature difference of the previous cycle.

[0079] Example 1:

[0080] This embodiment is based on a method for judging and correcting the error magnitude of heating system load prediction points:

[0081] 1. Project Overview. This example uses 32 days of operational data from a centralized heating system in a cold region, from February 12 to March 15, 2021. The prediction period is determined based on the control cycle of the heating station, resulting in a total of 123 prediction points during this period.

[0082] 2. The original prediction feature set FS is composed of the historical heat load of the previous three cycles, the secondary supply and return water temperature difference of the previous cycle, and the outdoor air temperature of the prediction cycle. A recursive multiple linear regression is used as the prediction model to predict the heat load of the heating station. The absolute percentage error (ARE) for each prediction point is calculated, and the distribution of ARE for the 123 prediction points is statistically analyzed as follows: Figure 3 As shown.

[0083] 3. Calculate the average daily outdoor air temperature over these 32 days. Then, subtract the previous day's average outdoor air temperature from the current day's average outdoor air temperature to calculate the daily change in average outdoor air temperature. The difference result is shown below. Figure 4As shown. Due to the thermal inertia of large-scale centralized heating systems and buildings, a time window width of 3 days is set, allowing this time window to slide across the difference results. A threshold is set. The time window is defined as 4℃. If there is a difference result with an absolute value greater than 4℃ within the time window, then the time window is determined to be a key point time window.

[0084] 4. Using this method, a total of 60 prediction points were identified for the periods from February 17th to February 27th and from March 4th to March 7th. The error distribution of these 60 prediction points is as follows: Figure 5 As shown, these 60 identified prediction points include all original prediction points with an ARE greater than 10%, and those with an ARE greater than 6% account for as much as 30%. The results demonstrate that this method can successfully identify points with large errors in advance, preparing for subsequent targeted error correction.

[0085] contrast Figure 3 and Figure 5 , Figure 3 This is the statistical result of the prediction error of 123 prediction points in the test set before correction. It can be seen that all errors are divided into six intervals: [0%, 2%), [2%, 4%), [4%, 6%), [6%, 8%), [8%, 10%), and [10%, +∞). More than 40% of the prediction points have errors less than 2%, while 9 prediction points have errors greater than 10%. It can be seen that the prediction accuracy of heat load is very high most of the time, but there are a small number of points with relatively large prediction errors.

[0086] Figure 5 The error distribution of 60 key points identified from the original 123 prediction points was obtained by using the sliding time window method on the difference results of outdoor temperature. The original 9 points with errors greater than 10% were successfully identified, and the proportion of points with ARE greater than 6% was as high as 30%. The results show that the method of the present invention can successfully identify points with large errors in advance.

[0087] 5. Based on FS, an error correction feature ε is introduced to obtain a new prediction feature set FS′. Recursive multiple linear regression is used as the prediction model to predict the heat load of the heating station. A comparison of the predicted heat load before and after correction with the ARE is shown in the figure below. Figure 6 As shown, compared to prediction using the original prediction feature set FS, especially for pre-identified key points, the prediction error after introducing the correction feature ε was effectively reduced, and the predicted heat load is closer to the actual heat load. Originally, 9 prediction points had an ARE greater than 10%, but after introducing the correction feature ε, only 4 prediction points had an ARE greater than 10%. Furthermore, the introduction of the correction feature ε did not significantly affect the prediction results for non-key points, indicating that introducing the correction only for the identified key points can more accurately reduce the error at these key points.

[0088] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for judging and correcting the error magnitude of load prediction points in a heating system, characterized in that, The method includes the following steps: To calculate the future daily average outdoor air temperature, obtain the average outdoor air temperature for each day, and then subtract the average outdoor air temperature of the previous day from the average outdoor air temperature of the current day. Obtain the difference results representing the changes in the daily average outdoor air temperature over the next few days; Set the threshold for daily average outdoor air temperature change. A time window with a width of p days is set, and this time window slides across the difference results; each time window contains p... If the condition satisfying equation (3) occurs within the time window, the time window is determined to be a key point time window, and the predicted point within the time window is the key point that has been identified. The original prediction feature set FS for heat load prediction is obtained by using feature selection. An error correction feature ε is introduced on the original prediction feature set FS to obtain a new prediction feature set FS′. FS=[X1,X2,…,X k ] (4) FS′=[X1,X2,…,X k ,e] (5) Where FS is the original prediction feature set; X is a single feature; k is the number of features; and ε is the error correction feature. Let q be the predicted heat load for the i-th period, in kW; i-1 The actual heat load value for the (i-1)th cycle, in kW; During the dynamic sliding judgment process, when the time window is determined to be a key point time window, the error correction characteristic is the predicted heat load value of the i-th cycle. Compared with the actual heat load q in the (i-1)th cycle i-1 The relative error is determined; otherwise, the error correction feature is set to 0. Using recursive multiple linear regression as the prediction model and FS′ as the prediction feature set, we can perform online dynamic prediction of future heat load.

2. The method for judging and correcting the error magnitude of the load prediction point of the heating system according to claim 1, characterized in that, The features in the original prediction feature set are: the historical heat load of the previous three cycles, the outdoor temperature of the prediction cycle, and the secondary temperature difference of the previous cycle, i.e. FS=[q i-1 ,q i-2 ,q i-3 ,t out,i ,Δt 2,i-1 ] Where, q i-1 ,q i-2 ,q i-3 For the historical heat load of the previous three cycles, t out,i To predict the outdoor temperature over a period of time, Δt 2,i-1 This is the second temperature difference of the previous cycle.

3. The method for judging and correcting the error magnitude of the load prediction point of the heating system according to claim 1, characterized in that, The predicted heat load value for the i-th period Obtained from the prediction model.

4. The method for judging and correcting the error magnitude of the load prediction point of the heating system according to claim 1, characterized in that, The average daily outdoor air temperature is obtained by formula (1). Among them, t out,n Let be the outdoor air temperature in the nth hour, in °C; The average daily outdoor air temperature is given in °C. Future changes in average daily outdoor air temperature Obtained from formula (2), in, The average outdoor air temperature on day j, in °C; The average outdoor air temperature on day j-1, in °C; Future outdoor air temperature is obtained from weather forecasts.

5. The method for judging and correcting the error magnitude of the load prediction point of the heating system according to claim 1, characterized in that, The k = 3-5; the The temperature range is 3℃-5℃; the cycle is the control cycle of the heating station, which is 5-8 hours.

Images