An optimized adjustment method for autonomous prediction of heat demand load

Abstract

The application discloses an optimization adjustment method for autonomous prediction of heating demand load, comprising the following steps: constructing a heating demand load time series by collecting heating system operation data and environmental data, and introducing a Hurst index to analyze the memory characteristics of the load time series to determine the evolution state of the heating load. Based on the load evolution state, a GM(1, N) grey prediction model is dynamically constructed to predict the future heating demand load. According to the predicted load result, the corresponding heating adjustment amount is generated, and the adjustment amount is converted into executable adjustment instructions to realize the forward-looking adjustment of the heating system operation. At the same time, the actual load result after the adjustment is executed is written back to update the load time series, and the Hurst index and the GM(1, N) grey prediction model parameters are recalculated to form a closed-loop operation process of mutual interaction between prediction and adjustment. The application realizes reliable prediction and on-demand adjustment of the heating demand load under the condition of limited data samples.

Description

Technical Field

[0001] This invention relates to the field of heating system operation control and data processing technology, and in particular to an optimized adjustment method for autonomous prediction of heating demand load. Background Technology

[0002] With the continuous expansion of urban centralized heating and the improvement of the automation level of heating systems, accurate prediction and operation regulation of heating demand load have become an important technical foundation for ensuring heating safety and improving energy efficiency. Currently, the operation and management of heating systems generally rely on historical load data or real-time monitoring data, and adjust heating parameters through empirical rules or fixed models to cope with changes in user heating demand.

[0003] Existing heating load forecasting and regulation technologies still have significant shortcomings. On the one hand, traditional load forecasting methods are mostly based on fixed time scales and static model parameters, making it difficult to reflect the long-term correlation and evolutionary characteristics of heating demand load over time. When the load is in a trend change or sudden change phase, the forecast results are prone to deviation, thus affecting the accuracy of regulation decisions. On the other hand, existing grey forecasting or statistical forecasting methods usually keep the parameters unchanged after the model is built, lacking a mechanism for identifying and responding to changes in load evolution, resulting in insufficient adaptability of the forecasting model under different load conditions. At the same time, existing heating regulation methods mostly adopt ex-post correction or passive response strategies, with regulation actions lagging behind load changes, easily leading to over-supply or under-supply, and making it difficult to form a closed-loop operation process linking forecasting and regulation.

[0004] Therefore, how to provide an optimized regulation method for autonomous prediction of heating demand load is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an optimized regulation method for autonomous prediction of heating demand load. This invention comprehensively utilizes time series memory analysis and grey system prediction methods. By introducing the Hurst exponent to identify the evolution characteristics of heating demand load, and adaptively constructing a GM(1,N) grey prediction model based on the load evolution state, it achieves the prediction of heating demand load. Based on this, heating regulation quantities are generated according to the predicted load and converted into regulation commands for proactive regulation of the heating system operation. Simultaneously, the actual load results after regulation execution are written back to form a prediction-regulation closed-loop operation. This method achieves synergistic optimization of load prediction and operation regulation under the condition of limited historical data, and has the advantages of high prediction accuracy, strong model adaptability, and good timeliness of heating regulation.

[0006] An optimization and adjustment method for autonomous prediction of heating demand load according to an embodiment of the present invention includes the following steps: Collect heating system operation data and environmental data, and perform time synchronization, missing data compensation and anomaly removal. Calculate heating demand load based on the processed data and generate load time series. A self-updating sliding time window is constructed based on the load time series, and the Hurst exponent of the load time series within the time window is calculated each time the time window is updated to generate a load memory criterion value. The load evolution state is determined based on the load memory criterion value. When the load evolution state changes, the reconstruction parameter generation process of the GM(1,N) grey prediction model is triggered to generate the GM(1,N) grey prediction model construction parameters. The GM(1,N) grey prediction model construction parameters include the modeling window length, the generation sequence transformation method and the parameter update rule. The parameters of the GM(1,N) grey prediction model are called, the data corresponding to the modeling window length are extracted from the load time series and the generation sequence transformation is performed to construct the GM(1,N) grey prediction model and output the heating demand load prediction sequence within the prediction period. Generate a heating regulation sequence based on the load forecast sequence and convert it into a regulation instruction set; The actual load sequence is obtained by executing the adjustment instruction set. The actual load sequence is written back to the load time series and the load memory criterion value is updated for the next round of load evolution state determination and GM(1,N) grey prediction model reconstruction. The prediction adjustment closed-loop operation results are output.

[0007] Optionally, the generation of the load time series includes: Collect heating system operation data and environmental data. The operation data includes supply water temperature, return water temperature, pipeline flow, heat source output and circulation equipment operation status. The environmental data includes outdoor temperature, wind speed and calendar time information. Record the corresponding collection timestamp for each data item to form the raw data set. A unified timeline is constructed based on the original dataset. Data from different sampling periods are resampled according to the unified timeline, so that the running data and the environmental data correspond one-to-one in the time dimension, resulting in a time-synchronized dataset. The time-synchronized data set is processed to identify missing data locations and fill in the missing data according to preset compensation rules to form a missing-compensated data set. Anomaly removal is performed on the data set after missing data compensation. Anomalies are identified based on the physical reasonable range and change constraints of the operational data and environmental data. Anomalies are then removed or replaced to form a cleaned data set. Based on the cleaned data set, the heating demand load value at each time point is calculated according to the correspondence between supply water temperature, return water temperature and pipeline flow rate. The calculated heating demand load values ​​are then arranged in a unified time axis order to generate a heating demand load time series.

[0008] Optionally, the generation of the load memory criterion value includes: Based on the heating demand load time series, the sliding time window length and time window update step size are set according to a unified time axis. At each update time, continuous load data corresponding to the time window length is extracted from the load time series to form the load subsequence within the current time window. The time window is then updated in a rolling manner according to the update step size as time progresses. After each sliding time window update, the average value of the load subsequence within the time window is calculated, and the average value is used to remove the mean of the load subsequence to obtain a centered load sequence for memory analysis. The cumulative deviation sequence is calculated in chronological order based on the centralized load sequence to characterize the offset characteristics of the load as it accumulates over time within the current time window; According to the preset multi-scale division rules, the load subsequence within the time window is divided into several sub-segments of different lengths. At each scale, the maximum and minimum values ​​of the cumulative deviation sequence within the corresponding sub-segment are calculated to obtain the segment range, and the dispersion index of the load subsequence within the corresponding sub-segment is calculated simultaneously. Based on the segment range and corresponding dispersion results obtained at different scales, a correspondence between scale and rescaled range is constructed. By linearly fitting the correspondence between scale and rescaled range, an index value reflecting the long-term correlation characteristics of load time series is obtained. The index value is used as the Hurst index corresponding to the current sliding time window, and associated with the time point corresponding to the end of the time window to form a load memory criterion value sequence that is updated with the sliding time window.

[0009] Optionally, the generation of parameters for the GM(1,N) grey prediction model includes: Based on the load memory criterion value sequence, the load memory criterion value corresponding to the current time window is read after each sliding time window update to generate the current state judgment input value; The current state determination input value is matched with the preset criterion interval set one by one, and the current load evolution state identifier is output according to the matching result. When the current state determination input value falls into the first criterion interval, a trend-continuous state identifier is output; when it falls into the second criterion interval, a random fluctuation state identifier is output; when it falls into the third criterion interval, a mean-regression state identifier is output. The current load evolution state identifier is associated with and stored with the end time point of the corresponding time window. Read the load evolution status identifier corresponding to the previous time window, compare the current load evolution status identifier with the previous load evolution status identifier, and generate a status change judgment result when the two are inconsistent. Use the status change judgment result as the reconstruction trigger signal of the GM(1,N) gray prediction model. When the GM(1,N) grey prediction model reconstruction trigger signal is in the trigger state, the GM(1,N) grey prediction model reconstruction parameter generation process is started. Based on the current load evolution state identifier, the modeling window length corresponding to the load evolution state identifier is selected from the preset parameter mapping table to generate the modeling window length. During the parameter reconstruction process, the generation sequence transformation method corresponding to the load evolution state identifier is selected from the preset transformation mapping table based on the current load evolution state identifier, and the generation sequence transformation method is generated. During the reconstruction of parameter generation, parameter update rules are generated by selecting the parameter update rules corresponding to the load evolution status identifier from the preset update mapping table based on the current load evolution status identifier. The modeling window length, the generation sequence transformation method, and the parameter update rule are combined to form the parameters for the GM(1,N) grey prediction model.

[0010] Optionally, the output of the heating demand load forecast sequence includes: Based on the GM(1,N) grey prediction model, the parameter reading modeling window length, the generation sequence transformation method and parameter update rules are constructed. According to the modeling window length, the load modeling subsequence is extracted from the heating demand load time series, and N-1 driving variable subsequences within the same time range as the load modeling subsequence are extracted simultaneously to form a modeling data group. Read the load evolution status identifier associated with the end time point of the modeling data group, perform state consistency segmentation processing within the modeling data group based on the load evolution status identifier, generate several continuous state consistency subsequence sets, and generate a corresponding segmented driving variable subsequence for each state consistency subsequence to form a segmented modeling data group. The generation process specified by the generation sequence transformation method is performed on each segment of the segmented modeling data group to generate the load generation sequence and the driving variable generation sequence of each segment respectively. Based on the load generation sequence of each segment, the background value sequence of each segment is constructed to form the segmented modeling input set. Based on the segmented modeling input set, a corresponding GM(1,N) sub-model modeling equation set is established for each consistent sub-sequence of states. The regression matrix and response vector corresponding to the GM(1,N) sub-model modeling equation set are assembled, and the parameter update rule is called to determine the parameter solution triggering condition of the GM(1,N) sub-model. When the parameter solution triggering condition is met, the least squares parameter solution is performed on the corresponding GM(1,N) sub-model to obtain the sub-model parameters, and the sub-model parameters are associated with the load evolution state identifier and time range of the GM(1,N) sub-model to generate a sub-model set; Based on the load evolution state identifier corresponding to the prediction starting point, the GM(1,N) sub-model with the same load evolution state identifier and the closest time range is retrieved from the sub-model set and selected as the prediction sub-model, and a load generation spatial prediction sequence is generated based on the prediction sub-model. During the generation of the load generation spatial forecast sequence, the parameter update rule is called to determine whether the sub-model parameter re-estimation of the forecast process is triggered. When triggered, the regression matrix and response vector are updated based on the latest piecewise modeling input set, and the sub-model parameters are updated to generate the updated load generation spatial forecast sequence. The updated load generation spatial prediction sequence is reverse generated to obtain the heating demand load prediction sequence within the prediction period, and the heating demand load prediction sequence is associated with and stored with the corresponding prediction time point.

[0011] Optionally, generating a heating regulation sequence based on the load forecast sequence and converting it into a regulation instruction set includes: Based on the heating demand load forecast sequence, the forecast load value is read point by point in the order of the forecast time, and the forecast load value is associated with the actual load value corresponding to the current operating state of the heating system to form the load regulation calculation input sequence; Based on the load regulation calculation input sequence, and according to the preset load-regulation mapping relationship, the heating regulation sequence is calculated point by point in time. The heating regulation sequence includes the heat source output regulation, water supply temperature regulation, and circulating equipment operating parameter regulation corresponding to the predicted load change direction and magnitude. The generated heating regulation sequence is processed according to equipment operation constraints. The heating regulation at each time point is superimposed with the current operating set value to obtain the corresponding target operating set value sequence. The target operating setpoint sequence is split according to equipment type, forming a heat source output setpoint sequence, a water supply temperature setpoint sequence, and a circulating equipment operating parameter setpoint sequence. Based on the sequence of target operating setpoints, the target operating setpoints and corresponding predicted time points are encapsulated in a unified instruction format to generate adjustment instruction entries containing time identifiers and operating setpoints. The adjustment instruction entries corresponding to each predicted time point are arranged in chronological order to form an adjustment instruction set.

[0012] Optionally, the output of the predictive regulation closed-loop operation results includes: After executing the set of adjustment instructions, the operating data after the adjustment is executed is collected according to the time identifier corresponding to the adjustment instruction item. The actual load sequence is calculated based on the operating data, where the actual load sequence is the sequence of observed heating demand load corresponding to each time identifier. The actual load sequence is aligned with the heating demand load forecast sequence according to the time identifier to generate a forecast-actual paired sequence, and a forecast error sequence is generated based on the forecast-actual paired sequence. The actual load sequence is written back to the heating demand load time sequence according to the time identifier. The actual load value is appended to the corresponding time point on a unified time axis to form an updated heating demand load time sequence. Based on the updated heating demand load time series, the load subsequence within the current time window is re-extracted according to the sliding time window length and the time window update step size, and the current load memory criterion value in the load memory criterion value sequence is regenerated based on the load subsequence within the current time window. The newly generated current load memory criterion value is used as the current state determination input value. The next round of load evolution state identifier is output, and the next round of load evolution state identifier is compared with the load evolution state identifier corresponding to the previous time window to generate the next round of state change determination result. When the state change determination result in the next round is that the change is valid, the GM(1,N) grey prediction model reconstruction parameter generation process is triggered, and the GM(1,N) grey prediction model construction parameters for the next round are output. The actual load sequence, prediction error sequence, updated heating demand load time sequence, current load memory criterion value, next round load evolution status identifier, and next round GM(1,N) grey prediction model are constructed and stored as parameters, and the prediction and regulation closed-loop operation results are output.

[0013] The beneficial effects of this invention are: This invention introduces a load time series memory analysis method, utilizing the Hurst exponent to characterize the long-term correlation and evolution of heating demand load. Based on the load evolution state, it adaptively constructs a GM(1,N) grey prediction model, achieving stable prediction of heating demand load under different operating conditions. This effectively reduces the prediction bias of traditional fixed models under load trend changes and abrupt changes in operating conditions. Furthermore, by constructing predictive sub-models based on the consistency of load evolution state and dynamically selecting a prediction model matching the current load state during the prediction process, the adaptability and robustness of the prediction results to actual operating conditions are improved.

[0014] This invention generates heating regulation quantities and forms a regulation instruction set based on predicted heating demand load, transforming the operation of the heating system from a passive response to a prediction-driven, forward-looking regulation. The actual load results after regulation execution are then written back to update the load time series and load memory criteria, forming a closed-loop operation mechanism of coordinated updating of prediction and regulation. Through these technical means, continuous adaptive optimization of the heating system's operating status is achieved under conditions of limited historical data and incomplete information, improving the stability, timeliness of regulation, and energy utilization efficiency of the heating system. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an optimized adjustment method for autonomous prediction of heating demand load proposed in this invention; Figure 2 This is a schematic diagram of the processing flow of the present invention, which adaptively constructs a GM(1,N) grey prediction model based on load evolution state and outputs a heating demand load prediction sequence. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 An optimized regulation method for autonomous prediction of heating demand load includes the following steps: Collect heating system operation data and environmental data, and perform time synchronization, missing data compensation and anomaly removal. Calculate heating demand load based on the processed data and generate load time series. A self-updating sliding time window is constructed based on the load time series, and the Hurst exponent of the load time series within the time window is calculated each time the time window is updated to generate a load memory criterion value. The load evolution state is determined based on the load memory criterion value. When the load evolution state changes, the reconstruction parameter generation process of the GM(1,N) grey prediction model is triggered to generate the GM(1,N) grey prediction model construction parameters. The GM(1,N) grey prediction model construction parameters include the modeling window length, the generation sequence transformation method and the parameter update rule. The parameters of the GM(1,N) grey prediction model are called, the data corresponding to the modeling window length are extracted from the load time series and the generation sequence transformation is performed to construct the GM(1,N) grey prediction model and output the heating demand load prediction sequence within the prediction period. Generate a heating regulation sequence based on the load forecast sequence and convert it into a regulation instruction set; The actual load sequence is obtained by executing the adjustment instruction set. The actual load sequence is written back to the load time series and the load memory criterion value is updated for the next round of load evolution state determination and GM(1,N) grey prediction model reconstruction. The prediction adjustment closed-loop operation results are output.

[0018] In this embodiment, the generation of the load time series includes: Collect heating system operation data and environmental data. The operation data includes supply water temperature, return water temperature, pipeline flow, heat source output and circulation equipment operation status. The environmental data includes outdoor temperature, wind speed and calendar time information. Record the corresponding collection timestamp for each data item to form the raw data set. A unified timeline is constructed based on the original dataset. Data from different sampling periods are resampled according to the unified timeline, so that the running data and the environmental data correspond one-to-one in the time dimension, resulting in a time-synchronized dataset. Resampling process constructs a common time axis by setting a unified sampling period. Data with a sampling period shorter than the unified sampling period is aggregated by segmented averaging, while data with a sampling period longer than the unified sampling period is completed by linear interpolation between adjacent time points, so that all data are aligned on the common time axis. The time-synchronized data set is processed to identify missing data locations and fill in the missing data according to preset compensation rules to form a missing-compensated data set. The location of missing data is identified by comparing a unified timeline with the timestamp sequences of each data point. When a certain time point exists on the unified timeline but the corresponding data value is empty or there is no valid sampling record, that time point is marked as the location of missing data. The preset compensation rules include tiered processing based on the duration of missing data, linear interpolation compensation using adjacent valid data for data with short consecutive missing durations, and compensation using the historical average under the same calendar conditions for data with long consecutive missing durations. The filling process assigns values ​​to the missing positions according to the preset compensation rules. When the interpolation method is used, the filling value is calculated based on the effective data adjacent to the missing point. When the historical average method is used, the historical statistical value under the corresponding calendar conditions is used as the filling value. Anomaly removal is performed on the data set after missing data compensation. Anomalies are identified based on the physical reasonable range and change constraints of the operational data and environmental data. Anomalies are then removed or replaced to form a cleaned data set. Anomaly removal is determined by comparing data values ​​with preset physical ranges and variation constraints. Data points that exceed reasonable ranges or show sudden changes are marked as anomalies and removed from the data sequence or replaced with adjacent valid data. The physical reasonable range is set according to the equipment design parameters to set the upper and lower limits of each data. The change constraint is to limit the data change amplitude between adjacent time moments. When the data exceeds the threshold or the adjacent change exceeds the preset amplitude, it is judged as abnormal data. When handling abnormal data, data points that are identified as abnormal are removed, and interpolated and replaced with adjacent valid data according to their position on the time axis, or replaced with the nearest valid data if interpolation is not possible. Based on the cleaned data set, the heating demand load value at each time point is calculated according to the correspondence between supply water temperature, return water temperature and pipeline flow rate. The calculated heating demand load values ​​are then arranged in a unified time axis order to generate a heating demand load time series. The heating demand load value is calculated at each time point by combining the supply water temperature, return water temperature and pipeline flow at the corresponding time. When the system directly provides the heat metering value, the metering value is used as the heating demand load at that time point.

[0019] In this embodiment, the generation of the load memory criterion value includes: Based on the heating demand load time series, the sliding time window length and time window update step size are set according to a unified time axis. At each update time, continuous load data corresponding to the time window length is extracted from the load time series to form the load subsequence within the current time window. The time window is then updated in a rolling manner according to the update step size as time progresses. The sliding time window length is determined by the time span covered by the most recent continuous load samples, and the time window update step size is determined by a fixed interval on a unified time axis; the length and step size are configured as preset values ​​during initialization and can remain unchanged for rolling updates according to the running cycle; Rolling updates discard the earliest load samples within each update time window and introduce the latest collected load samples, keeping the number of samples within the time window constant, thus enabling the time window to slide forward over time. After each sliding time window update, the average value of the load subsequence within the time window is calculated, and the average value is used to remove the mean of the load subsequence to obtain a centered load sequence for memory analysis. The average value is obtained by summing all load samples within the current time window and dividing by the number of samples. The number of samples is determined by the sliding time window length and the uniform sampling period. The resulting average value is used as the baseline value for this time window. The mean-reduction process subtracts the average value corresponding to the time window from each load sample within the time window to generate a centered load sequence, making the sequence centered at zero and retaining relative fluctuation information; The cumulative deviation sequence is calculated in chronological order based on the centralized load sequence to characterize the offset characteristics of the load as it accumulates over time within the current time window; According to the preset multi-scale division rules, the load subsequence within the time window is divided into several sub-segments of different lengths. At each scale, the maximum and minimum values ​​of the cumulative deviation sequence within the corresponding sub-segment are calculated to obtain the segment range, and the dispersion index of the load subsequence within the corresponding sub-segment is calculated simultaneously. The multi-scale partitioning rule divides the load subsequence within the time window continuously according to each scale by setting several segment scales of different lengths. The length of each scale is a different integer proportion of the length of the time window, which is used to form multi-scale analysis segments. Based on the segment range and corresponding dispersion results obtained at different scales, a correspondence between scale and rescaled range is constructed. By linearly fitting the correspondence between scale and rescaled range, an index value reflecting the long-term correlation characteristics of load time series is obtained. After obtaining the corresponding rescaled range values ​​at each scale, the length of each scale is paired with its corresponding rescaled range value to form a set of data relationships between scale and rescaled range. Linear fitting involves performing a logarithmic transformation on the set of relationships between the scale and the corresponding rescaled range, and then fitting a straight line using the least squares method to minimize the sum of squared deviations of each data point from the fitted line, thereby obtaining a stable fitting result. The index value is used as the Hurst index corresponding to the current sliding time window, and associated with the time point corresponding to the end of the time window to form a load memory criterion value sequence that is updated with the sliding time window.

[0020] In this embodiment, the generation of parameters for the GM(1,N) grey prediction model includes: Based on the load memory criterion value sequence, the load memory criterion value corresponding to the current time window is read after each sliding time window update to generate the current state judgment input value; The current state determination input value is matched with the preset criterion interval set one by one, and the current load evolution state identifier is output according to the matching result. When the current state determination input value falls into the first criterion interval, a trend-continuous state identifier is output; when it falls into the second criterion interval, a random fluctuation state identifier is output; when it falls into the third criterion interval, a mean-regression state identifier is output. The current load evolution state identifier is associated with and stored with the end time point of the corresponding time window. The preset criterion interval set is formed by dividing the range of load memory criterion values ​​into several continuous intervals. The boundaries of each interval are set during system initialization and correspond to different load evolution states, which are used for interval matching during state determination. Read the load evolution status identifier corresponding to the previous time window, compare the current load evolution status identifier with the previous load evolution status identifier, and generate a status change judgment result when the two are inconsistent. Use the status change judgment result as the reconstruction trigger signal of the GM(1,N) gray prediction model. When making comparisons, the load evolution status identifier corresponding to the current time window is compared with the load evolution status identifier corresponding to the previous time window one by one. When the two identifier values ​​are different, it is determined that the status has changed. When the GM(1,N) grey prediction model reconstruction trigger signal is in the trigger state, the GM(1,N) grey prediction model reconstruction parameter generation process is started. Based on the current load evolution state identifier, the modeling window length corresponding to the load evolution state identifier is selected from the preset parameter mapping table to generate the modeling window length. The process of reconstructing the parameters of the GM(1,N) grey prediction model refers to clearing the parameters of the previous model construction after detecting a change in the load evolution state, and entering the parameter generation process. Based on the current load evolution state, new modeling window lengths, generation sequence transformation methods, and parameter update rules are generated sequentially. The preset parameter mapping table is a state-parameter correspondence table established during system initialization. It is used to associate different load evolution states with the corresponding modeling window length values. Each load evolution state is configured with a unique corresponding modeling window length parameter in the table. During the parameter reconstruction process, the generation sequence transformation method corresponding to the load evolution state identifier is selected from the preset transformation mapping table based on the current load evolution state identifier, and the generation sequence transformation method is generated. The preset transformation mapping table is a state-transformation mode correspondence table established during system initialization. It is used to associate different load evolution states with the corresponding generation sequence transformation modes so as to quickly determine transformation parameters during model reconstruction. During the reconstruction of parameter generation, parameter update rules are generated by selecting the parameter update rules corresponding to the load evolution status identifier from the preset update mapping table based on the current load evolution status identifier. The preset update mapping table is a state-update rule correspondence table established during system initialization. It is used to associate different load evolution states with the corresponding parameter update rules so as to determine the parameter update method when the model is rebuilt. The modeling window length, the generation sequence transformation method, and the parameter update rule are combined to form the parameters for the GM(1,N) grey prediction model.

[0021] In this embodiment, the output of the heating demand load forecast sequence includes: Based on the GM(1,N) grey prediction model, the parameter reading modeling window length, the generation sequence transformation method and parameter update rules are constructed. According to the modeling window length, the load modeling subsequence is extracted from the heating demand load time series, and N-1 driving variable subsequences within the same time range as the load modeling subsequence are extracted simultaneously to form a modeling data group. Extracting load modeling subsequences refers to selecting continuous heating demand load data from the current time point backward on a unified time axis based on the modeling window length, and arranging them in chronological order to form a load data subsequence for model building. Read the load evolution status identifier associated with the end time point of the modeling data group, perform state consistency segmentation processing within the modeling data group based on the load evolution status identifier, generate several continuous state consistency subsequence sets, and generate a corresponding segmented driving variable subsequence for each state consistency subsequence to form a segmented modeling data group. Consistent state segmentation processing identifies the load evolution state identifiers corresponding to each time point in the load modeling subsequence, dividing continuous load data with the same state identifier into the same sub-segment, and data corresponding to different state identifiers into different sub-segments. The generation process specified by the generation sequence transformation method is performed on each segment of the segmented modeling data group to generate the load generation sequence and the driving variable generation sequence of each segment respectively. Based on the load generation sequence of each segment, the background value sequence of each segment is constructed to form the segmented modeling input set. The generation process specified by the generation sequence transformation method is to perform cumulative generation or specified generation operation on the load data and driving variable data in each consistent sub-segment according to the corresponding generation sequence transformation method, so as to form a generation sequence for modeling. Based on the segmented modeling input set, a corresponding GM(1,N) sub-model modeling equation set is established for each consistent sub-sequence of states. The regression matrix and response vector corresponding to the GM(1,N) sub-model modeling equation set are assembled, and the parameter update rule is called to determine the parameter solution triggering condition of the GM(1,N) sub-model. The GM(1,N) sub-model modeling equation set refers to the first-order multivariable grey differential equation discretization equation set established according to the grey system modeling rules for a single state consistent subsequence, based on the load generation sequence, the corresponding driving variable generation sequence, and the background value sequence of the subsequence. The regression matrix is ​​composed of the background value sequence and the generation sequence of driving variables corresponding to each time point, arranged in chronological order. The response vector is composed of the load generation sequence at the corresponding time point, and is used to describe the linear relationship between variables in the modeling equation. The parameter solving trigger condition for the GM(1,N) sub-model is determined by judging whether the sub-model is established for the first time, the load evolution state changes, or the preset update cycle is reached. When any of the conditions are met, the sub-model parameters are re-solved. When the parameter solution triggering condition is met, the least squares parameter solution is performed on the corresponding GM(1,N) sub-model to obtain the sub-model parameters, and the sub-model parameters are associated with the load evolution state identifier and time range of the GM(1,N) sub-model to generate a sub-model set; Least squares parameter solving refers to substituting the regression matrix and response vector into the linear equation system to calculate the model parameter values ​​that minimize the sum of squared predicted residuals, thereby obtaining the development coefficients of the GM(1,N) sub-model and the coefficients of each driving variable. Based on the load evolution state identifier corresponding to the prediction starting point, the GM(1,N) sub-model with the same load evolution state identifier and the closest time range is retrieved from the sub-model set and selected as the prediction sub-model, and a load generation spatial prediction sequence is generated based on the prediction sub-model. The GM(1,N) sub-model with the same load evolution state identifier and the closest time range refers to the sub-model selected from the sub-model set whose load evolution state identifier is the same as the current prediction state and whose modeling time interval end time is closest to the prediction start time. During the generation of the load generation spatial forecast sequence, the parameter update rule is called to determine whether the sub-model parameter re-estimation of the forecast process is triggered. When triggered, the regression matrix and response vector are updated based on the latest piecewise modeling input set, and the sub-model parameters are updated to generate the updated load generation spatial forecast sequence. The sub-model parameter re-estimation is triggered by detecting whether the load evolution state changes or the preset parameter update step number is reached during the prediction step. When either condition is met, the sub-model parameter re-estimation is triggered. Updating the regression matrix and response vector refers to reassembling the modeling data based on the latest consistent subsequences, and reconstructing the regression matrix and response vector accordingly. Then, the parameter solution is re-executed to update the parameters of the corresponding GM(1,N) sub-model. Perform inverse generation processing on the updated load generation spatial prediction sequence to obtain the heating demand load prediction sequence within the prediction period, and store the heating demand load prediction sequence in association with the corresponding prediction time point; Reverse generation processing refers to performing adjacent difference operations on the spatial prediction sequence of load generation in chronological order to obtain prediction values ​​with the same dimensions as the original load, and forming a load prediction sequence corresponding to the corresponding prediction time points.

[0022] In this embodiment, generating a heating regulation sequence based on the load prediction sequence and converting it into a regulation instruction set includes: Based on the heating demand load forecast sequence, the forecast load value is read point by point in the order of the forecast time, and the forecast load value is associated with the actual load value corresponding to the current operating state of the heating system to form the load regulation calculation input sequence; The actual load value corresponding to the current operating status is obtained by reading the heating demand load data that is closest to the prediction time point, and then aligning this load data with the system operating status data at the same time point to obtain the actual load value. Based on the load regulation calculation input sequence, and according to the preset load-regulation mapping relationship, the heating regulation sequence is calculated point by point in time. The heating regulation sequence includes the heat source output regulation, water supply temperature regulation, and circulating equipment operating parameter regulation corresponding to the predicted load change direction and magnitude. The preset load-adjustment mapping relationship is a correspondence rule established during system initialization. It is used to map load changes to adjustment values ​​of various operating parameters. Each adjustment value corresponds to the load change amplitude according to a fixed ratio or segmented rule. The generated heating regulation sequence is processed according to equipment operation constraints. The heating regulation at each time point is superimposed with the current operating set value to obtain the corresponding target operating set value sequence. Processing according to equipment operation constraints means limiting the amplitude and rate of change of the generated adjustment quantities, so that the target set values ​​of each operating parameter do not exceed the allowable range of the equipment and meet the requirements of continuous operation. Overlaying with the current operating setpoint means adding the calculated adjustment values ​​to the corresponding operating setpoints currently in effect for the equipment to generate the target operating setpoint for the next predicted time point, which is then used to generate subsequent adjustment instructions; The target operating setpoint sequence is split according to equipment type, forming a heat source output setpoint sequence, a water supply temperature setpoint sequence, and a circulating equipment operating parameter setpoint sequence. Segmenting by equipment type means classifying the target operating setpoints according to the controlled object and assigning them to the setpoint sequences corresponding to heat source equipment, water supply regulation equipment, and circulation equipment, which are then used to generate various regulation commands. Based on the sequence of target operating setpoints, the target operating setpoints and corresponding predicted time points are encapsulated in a unified instruction format to generate adjustment instruction entries containing time identifiers and operating setpoints. Segmenting by equipment type means classifying the target operating setpoints according to the controlled object and assigning them to the setpoint sequences corresponding to heat source equipment, water supply regulation equipment, and circulation equipment, which are then used to generate various regulation commands. The adjustment instruction entries corresponding to each predicted time point are arranged in chronological order to form an adjustment instruction set.

[0023] In this embodiment, the output of the predicted closed-loop operation result includes: After executing the set of adjustment instructions, the operating data after the adjustment is executed is collected according to the time identifier corresponding to the adjustment instruction item. The actual load sequence is calculated based on the operating data, where the actual load sequence is the sequence of observed heating demand load corresponding to each time identifier. The actual load sequence is aligned with the heating demand load forecast sequence according to the time identifier to generate a forecast-actual paired sequence, and a forecast error sequence is generated based on the forecast-actual paired sequence. The actual load sequence is written back to the heating demand load time sequence according to the time identifier. The actual load value is appended to the corresponding time point on a unified time axis to form an updated heating demand load time sequence. Based on the updated heating demand load time series, the load subsequence within the current time window is re-extracted according to the sliding time window length and the time window update step size, and the current load memory criterion value in the load memory criterion value sequence is regenerated based on the load subsequence within the current time window. The newly generated current load memory criterion value is used as the current state determination input value. The next round of load evolution state identifier is output, and the next round of load evolution state identifier is compared with the load evolution state identifier corresponding to the previous time window to generate the next round of state change determination result. When the state change determination result in the next round is that the change is valid, the GM(1,N) grey prediction model reconstruction parameter generation process is triggered, and the GM(1,N) grey prediction model construction parameters for the next round are output. The actual load sequence, prediction error sequence, updated heating demand load time sequence, current load memory criterion value, next round load evolution status identifier, and next round GM(1,N) grey prediction model are constructed and stored as parameters, and the prediction and regulation closed-loop operation results are output. Example

[0024] To verify the feasibility of this invention in practice, it was applied to a city-wide centralized heating system. This system serves multiple residential and public building heating units and uses a centralized heat source and pipeline distribution structure. While the system possesses basic operational data acquisition capabilities, it has long faced problems such as large fluctuations in heating demand load, insufficient forecast accuracy, and delayed adjustment response. Especially under conditions of frequent outdoor environmental changes and unstable user heating behavior, traditional operating methods relying on empirical rules or fixed prediction models struggle to reflect load change trends in a timely manner, often resulting in short-term oversupply or undersupply of heat, affecting system stability and causing energy waste.

[0025] In this application scenario, the operational and environmental data of the heating system are continuously collected first. Operational data includes supply water temperature, return water temperature, pipeline flow rate, heat source output, and the operating status of circulation equipment. Environmental data includes outdoor temperature, wind speed, and calendar time information. All data is mapped to the same timeline after collection and undergoes time synchronization, missing data compensation, and anomaly removal to form a stable and reliable data foundation. Based on this, the heating demand load is calculated according to the supply water temperature, return water temperature, and pipeline flow rate, and a heating demand load time series is constructed in chronological order.

[0026] During continuous system operation, a self-updating sliding time window is constructed using the load time series. Each time the time window is updated, a memory analysis is performed on the load series within the window to calculate the corresponding Hurst exponent. By analyzing the changes in the Hurst exponent, the system can identify whether the current load is in a different evolutionary state, such as a sustained trend, random fluctuation, or mean reversion. Unlike traditional methods, this embodiment does not assume that the load has a uniform change mechanism throughout the entire operating cycle; instead, it dynamically characterizes the load evolution features through load memory criteria.

[0027] When the system detects a change in load evolution, it automatically triggers the update process for the prediction model construction parameters. Based on the current load evolution state, it selects the corresponding modeling window length, generation sequence transformation method, and parameter update rules from the pre-configured parameter mapping relationship to generate the GM(1,N) grey prediction model construction parameters. On this basis, it performs state consistency segmentation processing on the load time series, dividing the load data within the modeling window into several sub-sequences with consistent evolution states, and constructs a corresponding GM(1,N) grey prediction sub-model for each sub-sequence. During the prediction process, based on the load evolution state corresponding to the prediction starting point, it selects the prediction sub-model with consistent state and the closest time range from the sub-model set to output the heating demand load prediction sequence.

[0028] In the actual operating data of this embodiment, it can be observed that the load exhibits obvious phased characteristics over a long period of time. For example, during periods of relatively stable environmental conditions, the Hurst exponent of the load time series remains stable between 0.65 and 0.75, indicating strong persistence in load changes; while during periods of greater environmental fluctuation, the Hurst exponent drops to close to 0.5, showing strong randomness in the load. Traditional fixed models use the same parameters for prediction in both types of phases, resulting in prediction lag during trend phases and amplified fluctuations during stochastic phases. This embodiment effectively avoids the above problems through state identification and sub-model selection.

[0029] In continuous operation tests, the method of this invention was compared with the original prediction and regulation methods. The results show that, under the same operating conditions, the average absolute error of the traditional method for load prediction is approximately 8% to 12%, and the error fluctuates significantly when the load changes drastically. However, after adopting the method of this invention, the average absolute error of prediction stabilizes in the range of 4% to 6%, and the error fluctuation is significantly reduced. Under partial load change conditions, the prediction error of the traditional method can exceed 15%, while the method of this invention, through state segmentation and sub-model matching, controls the maximum prediction error to within 8%.

[0030] After obtaining the predicted heating demand load sequence, this embodiment further generates a heating regulation sequence based on the prediction results. The system associates the predicted load with the actual load corresponding to the current operating state, and calculates the heat source output regulation, water supply temperature regulation, and circulating equipment operating parameter regulation according to the preset load-regulation mapping relationship. The regulation is subject to equipment operating constraints during the generation process and is superimposed with the current operating setpoint to form the target operating setpoint. Subsequently, the target operating setpoint is split according to equipment type and encapsulated into regulation instruction entries of a unified format, forming a regulation instruction set and issuing it for execution.

[0031] During the predictive regulation closed-loop operation, the system continuously collects actual operating data after regulation execution and calculates the actual load sequence. It then writes the actual load sequence back to update the load time series and recalculates the load memory criterion value for the next round of load evolution state determination and predictive model construction. Through this method, the system achieves closed-loop operation of prediction, regulation, and feedback.

[0032] In terms of actual operation, after adopting the method of this invention, the adjustment of heating system operating parameters is smoother, the adjustment frequency is reduced by about 20%, and the phenomena of over-supply and under-supply are significantly reduced. Under the same heating demand conditions, the fluctuation range of heat source output per unit time is reduced by about 15%, and the pipeline network operation is more stable. Comprehensive statistical results show that throughout the entire test period, the overall energy consumption of the system is reduced by about 6% to 9% compared with the original operation mode, while the heating comfort on the user side remains stable.

[0033] In summary, this embodiment fully demonstrates that by introducing load memory analysis and an adaptive construction method for the GM(1,N) grey prediction model based on load evolution state, and by forming a closed-loop operation mechanism between the prediction results and the heating regulation process, the problems of insufficient load prediction accuracy, poor model adaptability, and lag in regulation response in existing heating systems can be effectively solved. This ensures heating safety and stability while improving energy utilization efficiency.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An optimized regulation method for autonomous prediction of heating demand load, characterized in that, Includes the following steps: Collect heating system operation data and environmental data, and perform time synchronization, missing data compensation and anomaly removal. Calculate heating demand load based on the processed data and generate load time series. A self-updating sliding time window is constructed based on the load time series, and the Hurst exponent of the load time series within the time window is calculated each time the time window is updated to generate a load memory criterion value. The load evolution state is determined based on the load memory criterion value. When the load evolution state changes, the reconstruction parameter generation process of the GM(1,N) grey prediction model is triggered to generate the GM(1,N) grey prediction model construction parameters. The GM(1,N) grey prediction model construction parameters include the modeling window length, the generation sequence transformation method and the parameter update rule. The parameters of the GM(1,N) grey prediction model are called, the data corresponding to the modeling window length are extracted from the load time series and the generation sequence transformation is performed to construct the GM(1,N) grey prediction model and output the heating demand load prediction sequence within the prediction period. Generate a heating regulation sequence based on the load forecast sequence and convert it into a regulation instruction set; The actual load sequence is obtained by executing the adjustment instruction set. The actual load sequence is written back to the load time series and the load memory criterion value is updated for the next round of load evolution state determination and GM(1,N) grey prediction model reconstruction. The prediction adjustment closed-loop operation results are output.

2. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, The generation of load time series includes: Collect heating system operation data and environmental data. The operation data includes supply water temperature, return water temperature, pipeline flow, heat source output and circulation equipment operation status. The environmental data includes outdoor temperature, wind speed and calendar time information. Record the corresponding collection timestamp for each data item to form the raw data set. A unified timeline is constructed based on the original dataset. Data from different sampling periods are then resampled according to the unified timeline to obtain a time-synchronized dataset. The time-synchronized data set is processed to identify missing data locations and fill in the missing data according to preset compensation rules to form a missing-compensated data set. Anomaly removal is performed on the data set after missing data compensation. Anomalies are identified based on the physical reasonable range and change constraints of the operational data and environmental data. Anomalies are then removed or replaced to form a cleaned data set. Based on the cleaned data set, the heating demand load value at each time point is calculated according to the correspondence between supply water temperature, return water temperature and pipeline flow rate. The calculated heating demand load values ​​are then arranged in a unified time axis order to generate a heating demand load time series.

3. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, The generation of the load memory criterion value includes: Based on the heating demand load time series, the sliding time window length and time window update step size are set according to a unified time axis. At each update time, continuous load data corresponding to the time window length is extracted from the load time series to form the load subsequence within the current time window. The time window is then updated in a rolling manner according to the update step size as time progresses. After each sliding time window update, the average value of the load subsequence within the time window is calculated, and the average value is used to remove the mean of the load subsequence to obtain a centered load sequence for memory analysis. Calculate the cumulative deviation sequence based on the centered load sequence in chronological order; According to the preset multi-scale division rules, the load subsequence within the time window is divided into several sub-segments of different lengths. At each scale, the maximum and minimum values ​​of the cumulative deviation sequence within the corresponding sub-segment are calculated to obtain the segment range, and the dispersion index of the load subsequence within the corresponding sub-segment is calculated simultaneously. Based on the segment range and corresponding dispersion results obtained at different scales, a correspondence between scale and rescaled range is constructed. By linearly fitting the correspondence between scale and rescaled range, an index value reflecting the long-term correlation characteristics of load time series is obtained. The index value is used as the Hurst index corresponding to the current sliding time window, and associated with the time point corresponding to the end of the time window to form a load memory criterion value sequence that is updated with the sliding time window.

4. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, The generation of parameters for the GM(1,N) grey prediction model includes: Based on the load memory criterion value sequence, the load memory criterion value corresponding to the current time window is read after each sliding time window update to generate the current state judgment input value; The current state determination input value is matched with the preset criterion interval set one by one, and the current load evolution state identifier is output according to the matching result. When the current state determination input value falls into the first criterion interval, a trend-continuous state identifier is output; when it falls into the second criterion interval, a random fluctuation state identifier is output; when it falls into the third criterion interval, a mean-regression state identifier is output. The current load evolution state identifier is associated with and stored with the end time point of the corresponding time window. Read the load evolution status identifier corresponding to the previous time window, compare the current load evolution status identifier with the previous load evolution status identifier, and generate a status change judgment result when the two are inconsistent. Use the status change judgment result as the reconstruction trigger signal of the GM(1,N) gray prediction model. When the GM(1,N) grey prediction model reconstruction trigger signal is in the trigger state, the GM(1,N) grey prediction model reconstruction parameter generation process is started. Based on the current load evolution state identifier, the modeling window length corresponding to the load evolution state identifier is selected from the preset parameter mapping table to generate the modeling window length. During the parameter reconstruction process, the generation sequence transformation method corresponding to the load evolution state identifier is selected from the preset transformation mapping table based on the current load evolution state identifier, and the generation sequence transformation method is generated. During the reconstruction of parameter generation, parameter update rules are generated by selecting the parameter update rules corresponding to the load evolution status identifier from the preset update mapping table based on the current load evolution status identifier. The modeling window length, the generation sequence transformation method, and the parameter update rule are combined to form the parameters for the GM(1,N) grey prediction model.

5. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, The output of the heating demand load forecast sequence includes: Based on the GM(1,N) grey prediction model, the parameter reading modeling window length, the generation sequence transformation method and parameter update rules are constructed. According to the modeling window length, the load modeling subsequence is extracted from the heating demand load time series, and N-1 driving variable subsequences within the same time range as the load modeling subsequence are extracted simultaneously to form a modeling data group. Read the load evolution status identifier associated with the end time point of the modeling data group, perform state consistency segmentation processing within the modeling data group based on the load evolution status identifier, generate several continuous state consistency subsequence sets, and generate a corresponding segmented driving variable subsequence for each state consistency subsequence to form a segmented modeling data group. The generation process specified by the generation sequence transformation method is performed on each segment of the segmented modeling data group to generate the load generation sequence and the driving variable generation sequence of each segment respectively. Based on the load generation sequence of each segment, the background value sequence of each segment is constructed to form the segmented modeling input set. Based on the segmented modeling input set, a corresponding GM(1,N) sub-model modeling equation set is established for each consistent sub-sequence of states. The regression matrix and response vector corresponding to the GM(1,N) sub-model modeling equation set are assembled, and the parameter update rule is called to determine the parameter solution triggering condition of the GM(1,N) sub-model. When the parameter solution triggering condition is met, the least squares parameter solution is performed on the corresponding GM(1,N) sub-model to obtain the sub-model parameters, and the sub-model parameters are associated with the load evolution state identifier and time range of the GM(1,N) sub-model to generate a sub-model set; Based on the load evolution state identifier corresponding to the prediction starting point, the GM(1,N) sub-model with the same load evolution state identifier and the closest time range is retrieved from the sub-model set and selected as the prediction sub-model, and a load generation spatial prediction sequence is generated based on the prediction sub-model. During the generation of the load generation spatial forecast sequence, the parameter update rule is called to determine whether the sub-model parameter re-estimation of the forecast process is triggered. When triggered, the regression matrix and response vector are updated based on the latest piecewise modeling input set, and the sub-model parameters are updated to generate the updated load generation spatial forecast sequence. The updated load generation spatial prediction sequence is reverse generated to obtain the heating demand load prediction sequence within the prediction period, and the heating demand load prediction sequence is associated with and stored with the corresponding prediction time point.

6. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, Generating a heating regulation sequence based on the load forecast sequence and converting it into a regulation instruction set includes: Based on the heating demand load forecast sequence, the forecast load value is read point by point in the order of the forecast time, and the forecast load value is associated with the actual load value corresponding to the current operating state of the heating system to form the load regulation calculation input sequence; Based on the load regulation calculation input sequence, and according to the preset load-regulation mapping relationship, the heating regulation sequence is calculated point by point in time. The heating regulation sequence includes the heat source output regulation, water supply temperature regulation, and circulating equipment operating parameter regulation corresponding to the predicted load change direction and magnitude. The generated heating regulation sequence is processed according to equipment operation constraints. The heating regulation at each time point is superimposed with the current operating set value to obtain the corresponding target operating set value sequence. The target operating setpoint sequence is split according to equipment type, forming a heat source output setpoint sequence, a water supply temperature setpoint sequence, and a circulating equipment operating parameter setpoint sequence. Based on the sequence of target operating setpoints, the target operating setpoints and corresponding predicted time points are encapsulated in a unified instruction format to generate adjustment instruction entries containing time identifiers and operating setpoints. The adjustment instruction entries corresponding to each predicted time point are arranged in chronological order to form an adjustment instruction set.

7. The optimized adjustment method for autonomous prediction of heating demand load according to claim 1, characterized in that, The output of the predicted closed-loop operation results includes: After executing the set of adjustment instructions, the operating data after the adjustment is executed is collected according to the time identifier corresponding to the adjustment instruction item. The actual load sequence is calculated based on the operating data, where the actual load sequence is the sequence of observed heating demand load corresponding to each time identifier. The actual load sequence is aligned with the heating demand load forecast sequence according to the time identifier to generate a forecast-actual paired sequence, and a forecast error sequence is generated based on the forecast-actual paired sequence. The actual load sequence is written back to the heating demand load time sequence according to the time identifier. The actual load value is appended to the corresponding time point on a unified time axis to form an updated heating demand load time sequence. Based on the updated heating demand load time series, the load subsequence within the current time window is re-extracted according to the sliding time window length and the time window update step size, and the current load memory criterion value in the load memory criterion value sequence is regenerated based on the load subsequence within the current time window. The newly generated current load memory criterion value is used as the current state determination input value. The next round of load evolution state identifier is output, and the next round of load evolution state identifier is compared with the load evolution state identifier corresponding to the previous time window to generate the next round of state change determination result. When the state change determination result in the next round is that the change is valid, the GM(1,N) grey prediction model reconstruction parameter generation process is triggered, and the GM(1,N) grey prediction model construction parameters for the next round are output. The actual load sequence, prediction error sequence, updated heating demand load time sequence, current load memory criterion value, next round load evolution status identifier, and next round GM(1,N) grey prediction model are constructed and stored as parameters, and the prediction and regulation closed-loop operation results are output.

Images