Heat supply system heat load prediction method based on time sequence compensation and hybrid learning

Abstract

The invention provides a heat supply system thermal load prediction method based on time sequence compensation and hybrid learning, and the method comprises the steps: constructing a thermal load prediction input variable screening mechanism based on two-dimensional analysis, and achieving the scientific screening of input variables through the dual verification of a Pearson's correlation coefficient matrix and a significance test; a thermal load-temperature time sequence synchronization system based on dynamic time shift compensation is innovatively designed, and the problem of time sequence dislocation caused by instability of manual experience adjustment in a central heating system is solved; an ES-LSTM collaborative prediction architecture is provided, and a three-level prediction model of'trend decomposition-random learning-dynamic weighting 'is established. According to the method, statistics significance test and thermodynamic mechanism analysis are combined in a breakthrough manner, a dynamic compensation system with a time sequence self-correction capability is innovatively researched and developed, a hybrid model collaborative prediction mechanism is constructed, and a complete technical chain from data preprocessing to prediction model architecture is formed.

Description

Technical Field

[0001] The present invention relates to the field of smart heating, and more specifically, to a heat load prediction method for a heating system based on time series compensation and hybrid learning. Background Art

[0002] As an important component of urban infrastructure, the energy consumption of the centralized heating system accounts for a large proportion of the total energy consumption of buildings. With the acceleration of urbanization and the growth of energy demand, the operating efficiency and management level of the centralized heating system directly affect energy utilization efficiency and user comfort. In order to improve the operating efficiency of the centralized heating system and reduce energy waste, heat load forecasting has become a key technical means. Heat load forecasting is the basis for the operation and management of centralized heating systems. Accurate forecasts can help system operators adjust heating strategies in advance to avoid excessive energy consumption and waste. Short-term load forecasts (such as 1 hour, 2 hours, and 3 hours in advance) can be used to adjust the heating system in real time to adapt to changes in actual load and improve the system's operational safety and efficiency. Long-term load forecasts (such as 24 hours in advance) help managers formulate more reasonable operating strategies and achieve effective management of the entire heating system.

[0003] Throughout the development of heat load forecasting, modeling methods have evolved from traditional statistical models to artificial intelligence algorithms. Statistical methods such as ARMA and multivariate regression offer clear model structures and high computational efficiency, making them suitable for small-sample forecasting tasks. However, they struggle to capture nonlinear fluctuations and long-term dependencies in the data. In contrast, machine learning and deep learning models such as SVM and LSTM networks have improved forecasting accuracy through their powerful nonlinear modeling capabilities, demonstrating excellent fitting performance, particularly in the context of big data. Some researchers have also proposed ensemble learning and hybrid modeling strategies to achieve accurate forecasting across multiple time scales. However, many deep learning models ignore the significant temporal characteristics of heat load data, such as trends, seasonality, and randomness, during the modeling process, making them unable to fully reflect the actual heat load variation patterns. Furthermore, most methods assume strict alignment of historical time series data, ignoring the asynchrony of historical data in actual centralized heating systems due to factors such as human intervention and equipment failure. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention discloses a method for heat load forecasting for centralized heating systems based on time series compensation and hybrid learning. This method first introduces correlation analysis (using the Pearson correlation coefficient method) and significance level testing during the input data processing phase to ensure the statistical reliability of the relationship between the selected input variables and the heat load. Specifically, by proposing null and alternative hypotheses, constructing and calculating correlation statistics, and selecting appropriate significance levels for hypothesis testing, the method identifies key characteristic variables that have a significant impact on heat load forecasting.

[0005] During data preprocessing, this invention further integrates a dynamic time window mechanism and a bidirectional compensation algorithm to achieve time-series synchronization correction of historical data from the centralized heating system. The dynamic time window mechanism captures the time-varying lag between temperature changes and heat load responses in real time, while the bidirectional compensation algorithm aligns data based on the sensing results, ensuring the temporal consistency of the input data and thus eliminating the impact of historical data misalignment on model prediction accuracy.

[0006] In terms of predictive modeling, this paper combines an exponential smoothing model (ES) with a long short-term memory (LSTM) neural network to construct a hybrid prediction model. The ES equation is used to effectively capture the components of time series, such as trends and seasonality, while the LSTM neural network uses its deep learning capabilities to perform cross-learning and predict random components. This hybrid prediction model not only factors heat load data but also fully utilizes local and global characteristics to extract key information from the data, thereby improving prediction accuracy and providing reliable data support for the optimization and regulation of heating systems.

[0007] The technical means adopted by the present invention are as follows: a method for predicting heat load of a heating system based on time series compensation and hybrid learning, comprising:

[0008] S1: Construct a heat load prediction input variable screening mechanism based on dual-dimensional analysis, and use the Pearson correlation coefficient matrix and significance test method to double-verify the relationship between monitoring data and heat load in the heating system;

[0009] S2: Design a heat load temperature time series synchronization system based on dynamic time shift compensation to solve the time series data misalignment problem caused by the instability of manual experience adjustment in the centralized heating system;

[0010] S3: Design an ES-LSTM collaborative prediction architecture and establish a trend decomposition stochastic learning dynamic weighted three-level prediction model to predict the future short-term heat load of the centralized heating system.

[0011] Furthermore, the specific steps of step S1 include:

[0012] S11. Conduct data collection, covering multiple key data sources. The sources of heating data collection are wide-ranging, including thermal network monitoring data, meteorological data, and time data. Thermal network monitoring data is obtained by installing sensors at key nodes of the heating system. These sensors can monitor parameters such as supply and return water temperature, accumulated heat, and accumulated flow in real time. The circulation flow is obtained by subtracting the accumulated flow at two moments before and after, and the hourly heat, i.e., heat load, is obtained by subtracting the accumulated heat at two moments before and after, providing basic data for subsequent analysis and prediction. Meteorological data comes from professional meteorological monitoring stations and weather forecast platforms, including information such as outdoor temperature, wind speed, relative humidity, solar radiation, total cloud cover, and weather conditions. Time data includes date, day, and time, etc., and the time nodes of data collection are accurately recorded for subsequent time series analysis.

[0013] S12. The collected data needs to be preprocessed to ensure its quality and usability. During the data collection process, due to various reasons, the data may contain missing values ​​and outliers, which will affect the accuracy and reliability of the prediction model.

[0014] For missing data, interpolation is used to fill in the data. If the data before and after the moment is missing, the mean of the corresponding moments of the two adjacent days is used. This method can make up for the impact of missing data on analysis and prediction to a certain extent, and ensure the integrity of the data. For abnormal data, the box plot method is used to determine the outliers, and the interpolation method is used to replace the data. If there is long-term data missing or repeated due to abnormal collection process, such data will be directly eliminated. Through these preprocessing methods, the quality of the data can be effectively improved, providing a reliable data basis for subsequent heat load prediction. The interpolation calculation process is that if there are missing values ​​and outliers between two adjacent points (x1, y1) and (x2, y2), the interpolation is:

[0015]

[0016] S13. In order to improve the accuracy of heat load prediction and model efficiency, the correlation analysis method is used to screen out the input variables with the most significant impact. Correlation analysis is a method used to measure the degree of association between two or more variables, provided that there is a certain statistical relationship or probabilistic connection between the variables. By screening highly correlated variables, redundant information can be reduced and the computational efficiency and generalization ability of the model can be improved. In statistics, common correlation analysis methods are mainly used to evaluate the linear relationship between variables, among which the Pearson correlation coefficient is the most representative method. It measures the degree of linear correlation between two variables and is defined as the ratio of the covariance of the variables to the standard deviation, with a value range between -1 and 1. The value of the Pearson correlation coefficient has a clear statistical significance: when one variable increases, the other variable also increases accordingly, indicating that the two are positively correlated; when one variable increases, the other variable decreases, indicating that the two are inversely correlated; there is no linear correlation between the variables, that is, they are independent of each other. Through this method, the variables that have the greatest impact on heat load prediction can be effectively screened out, the input data can be optimized, and the stability and accuracy of the prediction model can be improved. The calculation formula of the Pearson correlation coefficient is as follows:

[0017]

[0018] The correlation analysis is shown in the table below:

[0019]

[0020] S14. After calculating the correlation coefficient, the significance level must be further calculated to ensure that the obtained variable relationship has a reliable statistical basis. The significance level calculation usually adopts the hypothesis testing method, which determines whether the observed correlation is caused by accidental factors under a given confidence level by solving the p-value. When the p-value is lower than the preset critical value (for example, 0.05), it can be considered that the correlation between the variables is statistically significant, thereby eliminating the influence of chance on the results. This process is particularly important for the heat load model, because only the variable relationship that has undergone rigorous significance testing can truly reflect the contribution of each factor to the change in heat load and provide solid data support for the accurate prediction and optimization control of the model. The hypothesis testing process is as follows:

[0021] (1) Propose the null hypothesis and alternative hypothesis:

[0022] H0:ρ=0

[0023] H1:ρ≠0

[0024] (2) Construct and calculate statistics, whose mathematical definition is as follows:

[0025]

[0026] (3) Select an appropriate significance level

[0027] The P value is the probability of an extreme or more extreme result if the null hypothesis is true. To test whether a value is significant, we specify a significance cutoff value, called the alpha value. In this paper, we use a 95% confidence interval and set the alpha value to 0.05.

[0028] Furthermore, in S2, a heat load-temperature timing synchronization system based on dynamic time shift compensation is designed to overcome the timing misalignment problem caused by the instability of manual experience adjustment in the centralized heating system.

[0029] The present invention adopts a timing synchronization technology that integrates a dynamic time window mechanism and a two-way compensation algorithm to eliminate the time series asynchrony of the historical data of the centralized historical heating system and avoid excessive errors in the prediction model caused by inputting incorrect time series. The dynamic window mechanism serves as the perception layer, responsible for capturing changes in the lead or lag rules. The two-way compensation algorithm is the execution layer, which can implement timing corrections based on the perception results. The dynamic time window mechanism ensures sensitivity to changes in the lag amount, while the two-way compensation algorithm converts the dynamic detection results into practical alignment operations through a safe and smooth correction strategy. The two together build a closed-loop compensation system from "perception" to "execution."

[0030] The core of the dynamic time window mechanism is to capture the time-varying lag relationship between temperature and heat load through sliding window cross-correlation analysis (SWCCA). Its implementation process includes the following key steps:

[0031] Set the window length W and the hysteresis search range τ∈[-5,5]. Choosing the window length requires a trade-off between sensitivity and stability: shorter windows are more responsive to sudden changes but are more susceptible to noise; longer windows are more resistant to noise but may mask short-term hysteresis changes. Therefore, it is necessary to find the optimal balance between sensitivity and stability in most scenarios.

[0032] To eliminate the dimensional differences and local fluctuation interference between the temperature and load series and enhance the robustness of the subsequent cross-correlation analysis, z-score normalization is first performed on the data within the window:

[0033]

[0034] Among them, μ T ,μ L is the mean of temperature and load in the window, σ T ,σ L is the standard deviation.

[0035] Calculate the cross-correlation function (CCF) of the normalized sequence within the window:

[0036]

[0037] By traversing the τ value to find the maximum point of the CCF absolute value, the optimal time shift τ is determined. opt To suppress the pseudo-peaks caused by noise, a local significance test is introduced. Only when the peak intensity exceeds the threshold θ = 0.7·max(|CCF(τ)|), is it determined to be a valid hysteresis. Otherwise, the result of the previous window is inherited.

[0038] The two-way compensation algorithm aims to convert the time-shift sequence {τ t} is transformed into a stable and secure timing alignment operation, the core of which includes time-shift sequence optimization and bidirectional correction strategy.

[0039] The original time shift {τ t It may contain high-frequency noise (such as disturbances caused by equipment startup and shutdown), and is smoothed using a weighted moving average:

[0040]

[0041]

[0042]

[0043] Where K is the half-window length. Gaussian filtering effectively suppresses random fluctuations while preserving trend changes, avoiding frequent jitter in compensation operations. To prevent excessive shifting of the time series, boundary constraints are added to ensure that correction operations always remain within the valid data range.

[0044] Assume that the original temperature sequence is T=[T1,T2,…,T N ], the load sequence is L=[L1,L2,…,L N ], the overall operation of the above dynamic timing synchronization method can be described as:

[0045]

[0046] Among them, g() represents the smoothing function of the time series, and CCF(τ) is the result of the sliding window cross-correlation analysis. This method integrates the dynamic time window mechanism and the two-way compensation algorithm to form a closed-loop system in the preprocessing stage. Through SVCCA, the lag law is captured in real time, dynamic fluctuations are covered, forward compensation is used to eliminate the lag effect, and backward compensation is used to correct the advance phenomenon. The problem of time series misalignment is solved uniformly, and Gaussian smoothing and index constraints are used to avoid data out of bounds and ensure the stability of the correction operation. In summary, this method can completely eliminate the time series misalignment caused by manual control and provide highly consistent input for the prediction model.

[0047] Furthermore, in S3, an ES-LSTM collaborative prediction architecture is proposed to establish a three-level prediction model of "trend decomposition-stochastic learning-dynamic weighting".

[0048] Traditional processing methods usually apply deseasonalization to time series during preprocessing to eliminate seasonal effects. Although this method can improve forecasting results to a certain extent, it has obvious shortcomings.

[0049] Classical methods use deseasonalization to decompose data, separating it from the forecast model. This can lead to incomplete and inconsistent forecasts. Furthermore, seasonal decomposition of data at the end of the time series often results in poor decomposition, thus affecting the accuracy of the overall forecast model. Furthermore, classic deseasonalization algorithms are not designed for deep learning and therefore have limitations when dealing with nonlinearities and complex interactions.

[0050] In contrast, the hybrid prediction model proposed by the present invention for the centralized heating system combines the advantages of classical statistical methods (exponential smoothing models) and deep learning algorithms, and regards seasonal processing as an integral part of the prediction model, rather than a preprocessing step independent of the prediction model, thereby improving the overall performance of the model. The principle of the smoothing exponential method is that exponential smoothing (ES) was proposed by Brown and later expanded by Holt and Winters. The main idea is to predict future observations based on previous observations, using a simple method to weight past observations in a time series to predict the future. This model is generally considered to be suitable for data without an expected upward or downward trend. The purpose is to estimate the true level. The estimate of this level is then used to predict future values. It is used in practice to predict trends and seasonality.

[0051] Because the heat load of the district heating system is always positive, a triple exponential smoothing (Holt-Winters) model with multiplicative seasonality was chosen. This triple exponential smoothing model decomposes the time series into trend, seasonal, and random components (residuals). This seasonal decomposition assumes that the components of the time series interact in a multiplicative manner.

[0052] In this model, non-seasonal (multiple heating seasons), single seasonality (independent heating seasons), and double seasonality (daily and hourly) models are used for district heating systems. The update formula for each case is as follows:

[0053] Non-seasonal model:

[0054] l t =αy t +(1-α)l t-1

[0055] Single-season model:

[0056] l t =αy t / s t +(1-α)l t-1

[0057] s t+K =βy t / l t +(1-β)s t

[0058] Dual Season Model:

[0059] l t =αy t / (s t u t )+(1-α)l t-1

[0060] s t+K =βy t / (l t u t )+(1-β)s t

[0061] u t+L =γy t / (l t s t )+(1-γ)u t

[0062] where y t is the value of the sequence at point t; l t , s t and u t are the trend, seasonality and second seasonality components respectively; K is the number of observations in each seasonal cycle; L is the number of observations in the smallest seasonal cycle. t and u t are always positive, and the smoothing coefficients α, β, γ∈[0,1].

[0063] Long Short-Term Memory (LSTM) is a derivative of neural networks. Compared to the original RNN, it introduces a gating mechanism to address the overall network model's issues of vanishing gradients, exploding gradients, and inability to handle long-term events caused by complex network layers. LSTM establishes a nonlinear relationship between output and historical input and embeds it into the neural network. Its core formula is as follows:

[0064] Oblivion: f t =W f ·[h t-1 ,x]+b f

[0065] Input: i t =tanh(W i ·[h t-1 ,x]+b i )

[0066] Unit status: c t =tanh(W c ·[h t-1 ,x]+b c )

[0067] Output: O t =W o [h t-1 ,x]+b o

[0068] Hidden layer: h t =o t ·tanh(c t )

[0069] A time series is the number of observations over a period of time. Therefore, it is necessary to analyze the coefficients of seasonal and horizontal time series. It consists of three layers: the first is the ES layer, which acts as a filter to capture and analyze the time series; the second is the LSTM layer, which is used to capture the randomness in the time series; and finally, the dense layer, whose task is to collect predictions from the first and second layers. In addition, the model also uses wind (speed, direction, and temperature) features to predict rainfall, as shown. Xt is the time-dependent input state, st refers to exponential smoothing and is time-dependent, and lt indicates that the LSTM is time-dependent.

[0070] In this system, the inputs to the neural network have been deseasonalized and normalized, assuming that the training loss function does not need to include normalization: it can just be a simple absolute difference between the target value and the predicted value. However, when observing its output, it can be seen that the model tends to have a positive bias, which is probably due to applying the squashing function log() to the NN's time series derived inputs and outputs. The system learns in logarithmic space, but the final prediction error is calculated in linear space. The system learns in logarithmic space, but the final prediction error is calculated in linear space. To overcome this, a quantile loss with a value of τ less than 0.5 is used. The quantile loss is defined as follows:

[0071]

[0072] In order to obtain an accurate and reliable prediction model, three evaluation criteria are determined, and their calculation formulas are shown below.

[0073] (1) Root mean square error (RMSE)

[0074]

[0075] (2) Mean absolute percentage error (MAPE)

[0076]

[0077] (3) Fitting coefficient (R2)

[0078]

[0079] Among them, Q j is the actual value; Q j ′ is the predicted value; n is the number of samples.

[0080] The evaluation indicators of the heat load prediction model include the root mean square error (RMSE), the mean percentage error (MAPE), and the fitting coefficient (R2). The closer the fitting coefficient is to 1, the smaller the MAPE and RMSE are, which means the smaller the prediction value error is and the higher the accuracy of the model is. Therefore, when evaluating the heat load prediction model, these indicators can be comprehensively considered to ensure the accuracy and reliability of the model.

[0081] Compared with the prior art, the present invention has the following advantages:

[0082] 1. The heat load prediction method for a heating system based on time series compensation and hybrid learning provided by the present invention adopts the Pearson correlation coefficient method combined with the significance level test to ensure the statistical correlation between the input variables and the heat load, improve the quality of the model input, and avoid the interference of irrelevant or weakly correlated variables on the prediction results.

[0083] 2. The heat load prediction method for a heating system based on time series compensation and hybrid learning provided by the present invention integrates a dynamic time window mechanism and a two-way compensation algorithm to eliminate the time series misalignment problem in the historical data of the heating system, improve the time consistency of the input data, and enable the prediction model to more accurately learn the true relationship between temperature changes and heat load responses.

[0084] 3. The heat load prediction method for a heating system based on time series compensation and hybrid learning provided by the present invention combines the exponential smoothing model (ES) with the long short-term memory neural network (LSTM), uses ES to extract trend and seasonal components, and uses LSTM to perform deep feature learning and random component prediction to achieve accurate modeling of heat load data.

[0085] 4. The present invention provides a method for predicting the heat load of a heating system based on time series compensation and hybrid learning. The hybrid prediction model can effectively decompose heat load data, make full use of local short-term characteristics and global long-term trends, improve data learning capabilities, and make the prediction results more stable and adaptable.

[0086] 5. The existing heat load prediction algorithm mainly performs overall prediction on time series. The present invention fully considers the influence of factors such as trend and seasonality in the time series of the centralized heating system, and improves the model's adaptability to complex time series characteristics.

[0087] Based on the above reasons, the present invention can be widely promoted in fields such as multi-unmanned boat formation path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0089] Figure 1 Flow chart of the method of the present invention.

[0090] Figure 2 This is a correlation heat map of an embodiment of the present invention.

[0091] Figure 3 This is an analysis of the time series synchronization method presented in an embodiment of the present invention.

[0092] Figure 4 This is an algorithm flow chart of the hybrid prediction method based on ES-DRNN shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0093] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0094] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0095] like Figure 1 As shown, the present invention provides a method for predicting heat load of a heating system based on time series compensation and hybrid learning, comprising:

[0096] S1: Construct a heat load prediction input variable screening mechanism based on dual-dimensional analysis, and use the Pearson correlation coefficient matrix and significance test method to double-verify the relationship between monitoring data and heat load in the heating system;

[0097] S2: Design a heat load temperature time series synchronization system based on dynamic time shift compensation to solve the time series data misalignment problem caused by the instability of manual experience adjustment in the centralized heating system;

[0098] S3: Design an ES-LSTM collaborative prediction architecture and establish a trend decomposition stochastic learning dynamic weighted three-level prediction model to predict the future short-term heat load of the centralized heating system.

[0099] In specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0100] A heat load prediction input variable screening mechanism based on dual-dimensional analysis was constructed. Through dual verification using the Pearson correlation coefficient matrix and significance test, scientific screening of input variables was achieved. The specific steps are as follows:

[0101] In order to improve the accuracy of heat load prediction and model efficiency, correlation analysis methods are used to screen out the most significantly influential input variables. Correlation analysis is a method used to measure the degree of association between two or more variables, presuming that there is a certain statistical relationship or probabilistic connection between the variables. By screening highly correlated variables, redundant information can be reduced, and the computational efficiency and generalization ability of the model can be improved. In statistics, common correlation analysis methods are mainly used to evaluate the linear relationship between variables, of which the Pearson correlation coefficient is the most representative method. It measures the degree of linear correlation between two variables and is defined as the ratio of the covariance of the variables to the standard deviation, with a value range between -1 and 1. The value of the Pearson correlation coefficient has clear statistical significance: when one variable increases, the other variable also increases accordingly, indicating a positive relationship between the two; when one variable increases and the other variable decreases, indicating an inverse relationship between the two; there is no linear correlation between the variables, that is, they are independent of each other. This method can effectively screen out the variables that have the greatest impact on heat load prediction, optimize input data, and improve the stability and accuracy of the prediction model. The calculation formula for the Pearson correlation coefficient is as follows:

[0102]

[0103] Taking the hourly heat consumption per unit area of ​​the heating station in the centralized heating system as the dependent variable, the influence of weather and operating parameters such as outdoor temperature, outdoor solar radiation, and primary water supply temperature on the hourly heat consumption per unit area is studied. Each influencing factor is taken as the variable x, and the hourly heat consumption per unit area is taken as the variable y. The calculation results are shown in Figure 2 .

[0104] In specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

[0105] S2. Design a heat load-temperature time sequence synchronization system based on dynamic time shift compensation to overcome the time sequence misalignment problem caused by the instability of manual experience adjustment in the centralized heating system.

[0106] Set the window length W and the hysteresis search range τ∈[-5,5]. Choosing the window length requires a trade-off between sensitivity and stability: shorter windows are more responsive to sudden changes but are more susceptible to noise; longer windows are more resistant to noise but may mask short-term hysteresis changes. Therefore, it is necessary to find the optimal balance between sensitivity and stability in most scenarios.

[0107] To eliminate the dimensional differences and local fluctuation interference between the temperature and load series and enhance the robustness of the subsequent cross-correlation analysis, z-score normalization is first performed on the data within the window:

[0108]

[0109] Among them, μ T ,μ L is the mean of temperature and load in the window, σ T ,σ L is the standard deviation.

[0110] Calculate the cross-correlation function (CCF) of the normalized sequence within the window:

[0111]

[0112] By traversing the τ value to find the maximum point of the CCF absolute value, the optimal time shift τ is determined. opt To suppress the pseudo-peaks caused by noise, a local significance test is introduced. Only when the peak intensity exceeds the threshold θ = 0.7·max(|CCF(τ)|), is it determined to be a valid hysteresis. Otherwise, the result of the previous window is inherited.

[0113] The two-way compensation algorithm aims to convert the time-shift sequence {τ t} is transformed into a stable and secure timing alignment operation, the core of which includes time-shift sequence optimization and bidirectional correction strategy.

[0114] The original time shift {τ t It may contain high-frequency noise (such as disturbances caused by equipment startup and shutdown), and is smoothed using a weighted moving average:

[0115]

[0116] Where K is the half-window length. Gaussian filtering effectively suppresses random fluctuations while preserving trend changes, avoiding frequent jitter in compensation operations. To prevent excessive shifting of the time series, boundary constraints are added to ensure that correction operations always remain within the valid data range.

[0117] Assume that the original temperature sequence is T=[T1,T2,…,T N ], the load sequence is L=[L1,L2,…,L N ], the overall operation of the above dynamic timing synchronization method can be described as:

[0118]

[0119] Among them, g() represents the smoothing function of the time series, and CCF(τ) is the result of sliding window cross-correlation analysis.

[0120] Figure 3 The computational results of the timing synchronization technology integrating the dynamic time window mechanism and the bidirectional compensation algorithm are demonstrated.

[0121] In specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

[0122] S3, propose ES-LSTM collaborative prediction architecture, such as Figure 4 The algorithm flow chart of the hybrid prediction method based on ES-DRNN is introduced, and a three-level prediction model of "trend decomposition-random learning-dynamic weighting" is established as follows:

[0123] The Holt-Winters model, based on the multiplicative seasonality, decomposes the time series into trend, seasonal, and random components (residuals). This seasonal decomposition assumes that the components of the time series interact in a multiplicative manner.

[0124] In this model, non-seasonal (multiple heating seasons), single seasonality (independent heating seasons), and double seasonality (daily and hourly) models are used for district heating systems. The update formula for each case is as follows:

[0125] Non-seasonal model:

[0126] l t =αy t +(1-α)l t-1

[0127] Single-season model:

[0128] l t =αy t / s t +(1-α)l t-1

[0129] s t+K =βy t / l t +(1-β)s t

[0130] Dual Season Model:

[0131] l t =αy t / (s t u t )+(1-α)l t-1

[0132] s t+K =βy t / (l t u t )+(1-β)s t

[0133] u t+L =γy t / (l t st )+(1-γ)u t

[0134] where y t is the value of the sequence at point t; l t , s t and u t are the trend, seasonality and second seasonality components respectively; K is the number of observations in each seasonal cycle; L is the number of observations in the smallest seasonal cycle. t and u t are always positive, and the smoothing coefficients α, β, γ∈[0,1].

[0135] Long Short-Term Memory (LSTM) is a derivative of neural networks. Compared to the original RNN, it introduces a gating mechanism to address the overall network model's issues of vanishing gradients, exploding gradients, and inability to handle long-term events caused by complex network layers. LSTM establishes a nonlinear relationship between output and historical input and embeds it into the neural network. Its core formula is as follows:

[0136] Oblivion: f t =W f ·[h t-1 ,x]+b f

[0137] Input: i t =tanh(W i ·[h t-1 ,x]+b i )

[0138] Unit status: c t =tanh(W c ·[h t-1 ,x]+b c )

[0139] Output: O t =W o [h t-1 ,x]+b o

[0140] Hidden layer: h t =o t ·tanh(c t )

[0141] A time series is the number of observations over a period of time. Therefore, it is necessary to analyze the coefficients of seasonal and horizontal time series. It consists of three layers: the first is the ES layer, which acts as a filter to capture and analyze the time series; the second is the LSTM layer, which is used to capture the randomness in the time series; and finally, the dense layer, whose task is to collect predictions from the first and second layers. In addition, the model also uses wind (speed, direction, and temperature) features to predict rainfall, as shown. Xt is the time-dependent input state, st refers to exponential smoothing and is time-dependent, and lt indicates that the LSTM is time-dependent.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 method for heat load forecasting of a central heating system based on time series compensation and hybrid learning, the characteristics of which include: S1: Construct a heat load prediction input variable screening mechanism based on dual-dimensional analysis, and use the Pearson correlation coefficient matrix and significance test method to double-verify the relationship between monitoring data and heat load in the heating system; S2: Design a heat load temperature time series synchronization system based on dynamic time shift compensation to solve the time series data misalignment problem caused by the instability of manual experience adjustment in the centralized heating system; S3: Design an ES-LSTM collaborative prediction architecture and establish a trend decomposition stochastic learning dynamic weighted three-level prediction model to predict the future short-term heat load of the centralized heating system.

2. The method for heat load prediction of a heating system based on time series compensation and hybrid learning according to claim 1 is characterized in that: S11: Acquire heat pipe network monitoring data through node sensors in the heating system. The monitoring data includes flow, heat, temperature, relevant meteorological data, and time data. The circulation flow is obtained by subtracting the accumulated flow at two moments, and the hourly heat is obtained by subtracting the accumulated heat at two moments. The meteorological data includes outdoor temperature, wind speed, relative humidity, solar radiation, total cloud cover, and weather condition information. The time data includes date, day, and time information. S12: preprocess the acquired monitoring data, fill in missing data using interpolation, identify outliers using box plot method, and replace outliers using interpolation; S13: For the pre-processed monitoring data, use the correlation analysis method to screen out the input variables that have the most significant impact on the heat load X. By calculating the Pearson correlation coefficient r between the current variable Y and the heat load X, the variables with the greatest influence on the heat load prediction are screened out: When |r| is between 0-0.3, it indicates almost no correlation, and when |r| is between 0.3-0.5, it indicates weak correlation. When |r| is between 0.5-0.8, it indicates low correlation, and when |r| is between 0.8-1.0, it indicates high correlation; S14: After calculating the Pearson correlation coefficient between different variables and heat load, the significance level is calculated. The hypothesis testing process is as follows: Propose the null hypothesis and alternative hypothesis: H0:ρ=0 H1:ρ≠0 Construct and calculate statistics, whose mathematical definitions are as follows: Select an appropriate significance level (P value). The P value is the probability of an extreme or more extreme result if the null hypothesis is true. Set a significance threshold. For the establishment of the alpha value, if the study is based on a 95% confidence interval, set the alpha value to 0.

05. S15: Based on the Pearson correlation coefficient r and significance level P calculated in S13 and S14, data with a Pearson correlation coefficient |r| > 0.65 and a significance level P < 0.05 are selected from the data as input variables of the prediction model.

3. The method for heat load prediction of a heating system based on time series compensation and hybrid learning according to claim 1, characterized in that: S21: For the historical time series heat load data, the dynamic time window mechanism is used to capture the time-varying lag relationship between outdoor temperature and heat load through sliding window cross-correlation analysis. First, the window length W and the lag search range τ∈[-5,5] are set for the dynamic window; S22: To eliminate the dimensional differences and local fluctuation interference between the outdoor temperature and load series and enhance the robustness of the correlation analysis, z-score normalization is performed on the data within the window: Among them, μ T ,μ L is the mean of outdoor temperature and load in the window, σ T ,σ L is the standard deviation; S23: Calculate the cross-correlation function CCF of the normalized sequence within the window length: By traversing the τ value to find the maximum point of the absolute value of CCF, the optimal time shift {τ t To suppress the pseudo-peaks caused by noise, a local significance test is introduced. Only when the peak intensity exceeds the threshold θ = 0.7 max(|CCF(τ)|), is it determined to be a valid hysteresis. Otherwise, the result of the previous window is inherited. S23: Use the two-way compensation algorithm to convert the time shift sequence {τ t } is converted into a stable and safe time series, and the original time shift {τ t } for smoothing: Where K is the half window length.

4. The method for heat load prediction of a heating system based on time series compensation and hybrid learning according to claim 1, characterized in that: S31: Use the smoothing exponential method to identify trend terms, seasonal components, and random residuals in the historical heat load data of the central heating system. Select the cubic exponential smoothing model with multiplicative seasonality. The update formula for each case is as follows: Non-seasonal model: l t ＝αy t +(1-α)l t-1 Single-season model: l t =αy t / s t +(1-α)l t-1 s t+K =βy t / l t +(1-β)s t Dual Season Model: l t =αy t / (s t the t )+(1-α)l t-1 s t+K =βy t / (l t u t )+(1-β)s t you t+L =γy t / (l t s t )+(1-γ)u t where y t is the value of the sequence at point t; l t 、s t and u t are the trend, seasonality and second seasonality components respectively; K is the number of observations in each seasonal cycle; L is the number of observations in the smallest unit seasonal cycle, where s t and u t is always positive, and the smoothing coefficients α, β, γ∈[0,1]; S32: Build an LSTM network based on a gating mechanism. LSTM establishes a nonlinear relationship between output and historical input and embeds it into the neural network. Its core formula is as follows: Oblivion: f t =W f ·[h t-1 ,x]+b f Input: i t =tanh(W i ·[h t-1 ,x]+b i ) Unit status: c t =tanh(W c ·[h t-1 ,x]+b c ) Output: O t =W o [h t-1 ,x]+b o Hidden layer: h t =o t ·tanh(c t ) The LSTM network is used to capture the nonlinear characteristics and long-distance dependencies of the heat load sequence and establish a mapping between historical input and future output. S33: The trend and seasonal terms decomposed from the exponential smoothing model are used as input features of the LSTM network. By sharing the LSTM network weights, the common patterns of different heat load series are extracted. Independent smoothing coefficients and seasonal cycle parameters are set for each series to preserve its unique seasonal structure. S34: Based on the trained ES-LSTM model, the deseasonalized data at the current moment and the historical trend seasonal characteristics are input to predict the future heat load value, and the deseasonalized predicted value output by LSTM is converted to The seasonal component at the corresponding time Multiply and restore to the actual heat load forecast value: S35: Use root mean square error RMSE, mean absolute percentage error MAPE, and fitting coefficient R 2 Evaluate the performance of the prediction model. The closer the fitting coefficient is to 1, the smaller the MAPE and RMSE are, which means the smaller the prediction error is and the higher the accuracy of the model is.

Images