Intelligent prediction method and system for heat supply load of heat station

Abstract

The application discloses a kind of heat station heat supply load intelligent prediction method and system, it is related to load intelligent prediction field, first to original heat station telemetry and weather data are fine and flow pre-processing, generate heat load prediction microcosmic sequence input, macroscopic sequence input and macroscopic context input.Subsequently, feature extraction is carried out for each scale input, to obtain respective embedding vector.The key is to introduce dynamic attention fusion gate, this module can intelligently learn and dynamically allocate the weight of each scale embedding vector, effectively solve the conflict and priority problem between different time scale signals, avoid the limitations in traditional method Hard coding, and expand the model receptive field.Finally, the multi-scale information after fusion is input into regressor for prediction, significantly improve the accuracy and robustness of prediction, overcome the defects that existing scheme cannot effectively handle multi-scale dynamic changes.

Description

Technical Field

[0001] This application relates to the field of intelligent load forecasting, and more specifically, to a method and system for intelligent forecasting of heating load in a heating station. Background Technology

[0002] In urban centralized heating systems, especially in cities where heavy industries such as oil extraction and processing are the economic pillars, efficient energy utilization and stable system operation are core objectives. To promote energy conservation, emission reduction, and a circular economy, heating systems in these cities often utilize waste heat generated during industrial processes such as oil refining as a significant heat source. While this model greatly improves overall energy utilization, it also presents new challenges for the scheduling and management of the heating system. The heating load of a heating station depends not only on traditional meteorological factors and user heating behavior but also on the production plans and operational fluctuations of upstream oil refining enterprises. Furthermore, the supply of heat sources themselves exhibits dynamic characteristics. Therefore, constructing an intelligent solution capable of accurately predicting heating load is not only a response to downstream user demand but also a crucial step in achieving synergistic optimization between the heating system and industrial production systems, ensuring a balance in energy supply.

[0003] However, when faced with the complex task of forecasting heat loads that integrates industrial production rhythms with residential heating patterns, existing heating station load forecasting schemes generally exhibit inherent limitations. Traditional forecasting methods, such as those based on statistical models or shallow machine learning models, struggle to fully capture the inherent multi-timescale dynamic characteristics of heat load data. Changes in heat load are not the result of a single scale, but rather a comprehensive reflection of signals superimposed at multiple timescales, including minute-level instantaneous fluctuations, hourly periodic patterns, and daily or even weekly trends. Existing models often employ simple feature concatenation or fixed-window moving averages—hard-coding methods—when processing this multi-scale information. This not only easily leads to feature redundancy, but more importantly, it lacks the ability to dynamically determine the importance of signals at different timescales, failing to effectively resolve conflicts and prioritization issues between signals of different scales. In addition, many models only focus on recent data, and their sensing field is limited, making it difficult to capture more macroscopic and periodic load patterns. This results in poor prediction results during periods of drastic load changes or special periods (such as peak heating times in the morning and evening, holidays, or sudden changes in upstream industrial heat sources). They are also easily affected by local noise, which can lead to energy waste or insufficient heating.

[0004] In view of the above challenges, an optimized intelligent prediction scheme for heating load of heating stations is desired, aiming to fundamentally improve the baseline accuracy of prediction models. Summary of the Invention

[0005] Considering the limitations mentioned above in application, according to one aspect of this application, a method for intelligent prediction of heating load at a heat station is provided, comprising: acquiring raw telemetry data streams and raw meteorological data streams from the heat station; performing split preprocessing on the raw telemetry data streams and raw meteorological data streams to obtain micro-scale sequence input, meso-scale sequence input, and macro-scale context input for heat load prediction; performing multi-timescale feature extraction on the micro-scale sequence input, meso-scale sequence input, and macro-scale context input for heat load prediction to obtain micro-scale embedding vector, meso-scale embedding vector, and macro-scale embedding vector for heat load prediction; inputting the micro-scale embedding vector, meso-scale embedding vector, and macro-scale embedding vector for heat load prediction into a dynamic attention fusion gate to obtain a multi-scale fusion embedding vector for heat load prediction; and inputting the multi-scale fusion embedding vector for heat load prediction into a prediction head based on a regressor to obtain a final predicted heat load value.

[0006] According to another aspect of this application, a smart forecasting system for heating load at a heat station is provided, comprising: a heat station weather data acquisition module for acquiring raw heat station telemetry data streams and raw meteorological data streams; a data splitting and preprocessing module for splitting and preprocessing the raw heat station telemetry data streams and the raw meteorological data streams to obtain a micro-sequence input for heat load forecasting, a meso-sequence input for heat load forecasting, and a macro-context input for heat load forecasting; and a multi-scale heat load forecasting encoding module for encoding the micro-sequence input for heat load forecasting, the meso-sequence input for heat load forecasting, and the macro-context input for heat load forecasting. The context input is used to extract features at multiple time scales to obtain microscale embedding vectors, mesoscale embedding vectors, and macroscale embedding vectors for heat load prediction. A multiscale fusion module for heat load prediction is used to input the microscale embedding vectors, mesoscale embedding vectors, and macroscale embedding vectors for heat load prediction into a dynamic attention fusion gate to obtain a multiscale fused embedding vector for heat load prediction. A heat load prediction module is used to input the multiscale fused embedding vector for heat load prediction into a prediction head based on a regressor to obtain the final heat load prediction value.

[0007] Compared with existing technologies, this application provides a method and system for intelligent prediction of heating load at heating stations. First, it performs refined preprocessing of the original telemetry and meteorological data from the heating station, generating micro-sequence input, meso-sequence input, and macro-context input for heating load prediction. This captures information at different time granularities at the data level, overcoming the shortcomings of traditional methods in processing multi-scale data. Subsequently, feature extraction is performed on the inputs at each scale to obtain their respective embedding vectors. The key lies in introducing a dynamic attention fusion gate. This module can intelligently learn and dynamically allocate the weights of the embedding vectors at each scale, effectively resolving conflicts and priority issues between signals at different time scales, avoiding the limitations of hard coding in traditional methods, and expanding the model's receptive field. Finally, the fused multi-scale information is input into a regressor for prediction, significantly improving the accuracy and robustness of the prediction and overcoming the shortcomings of existing schemes in effectively handling multi-scale dynamic changes. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a method for intelligent prediction of heating load at a heat station according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow of the intelligent prediction method for heating load of a heat station according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step S3 in the intelligent prediction method for heating load of a heating station according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of step S4 in the intelligent prediction method for heating load of a heating station according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of a heating station heating load intelligent prediction system according to an embodiment of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] To address the limitations of existing technologies, this application proposes an intelligent prediction method for heating load of heating stations. Figure 1 This is a flowchart of a method for intelligent prediction of heating load at a heat station according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the intelligent prediction method for heating load of a heat station according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent prediction method for heating load of a heat station according to an embodiment of this application includes: S1, acquiring the original heat station telemetry data stream and the original meteorological data stream; S2, performing split preprocessing on the original heat station telemetry data stream and the original meteorological data stream to obtain a micro-scale sequence input, a meso-scale sequence input, and a macro-scale context input for heat load prediction; S3, performing multi-time-scale feature extraction on the micro-scale sequence input, the meso-scale sequence input, and the macro-scale context input for heat load prediction to obtain a micro-scale embedding vector, a meso-scale embedding vector, and a macro-scale embedding vector for heat load prediction; S4, inputting the micro-scale embedding vector, the meso-scale embedding vector, and the macro-scale embedding vector for heat load prediction into a dynamic attention fusion gate to obtain a multi-scale fusion embedding vector for heat load prediction; S5, inputting the multi-scale fusion embedding vector for heat load prediction into a prediction head based on a regressor to obtain the final predicted heat load value.

[0016] In step S1, the original telemetry data stream and original meteorological data stream of the heating station are acquired. It should be understood that accurately predicting the heating load of heating stations is crucial for ensuring the efficient and economical operation of urban heating systems. Load changes are influenced by meteorological factors such as outdoor temperature, wind speed, and sunshine, as well as non-meteorological factors such as user behavior, building type, and holidays, exhibiting complex multi-timescale fluctuation characteristics. Existing prediction methods often use feature splicing or fixed-window moving averages to process multi-source information, making it difficult to dynamically identify the importance of signals at different time scales. This easily leads to feature redundancy and the curse of dimensionality, and the model's receptive field is limited, failing to effectively capture long-term periodic patterns and exhibiting poor noise resistance. Therefore, the prerequisite for building an intelligent prediction model is to comprehensively acquire telemetry and meteorological data from the heating station. Based on this, deep extraction and fusion of multi-timescale features can be achieved, overcoming the limitations of traditional methods in data dimension and time scale modeling, thereby improving prediction accuracy and solving the prediction bias problem caused by incomplete data or insufficient utilization of multi-scale information in current technologies.

[0017] In an optional implementation, step S1 is achieved through the following process: acquiring the raw telemetry data stream from the heating station. This data stream is acquired in real time from the heating station's SCADA system or local database. Considering the rapid changes in heat load, the data stream is acquired at a high frequency, set to a minute-level frequency, for example, once per minute. The minute-level frequency is chosen to capture rapid changes in heating load while avoiding excessive noise that might arise from second-level data, ensuring the validity and usability of the data. In addition to the heating load value used for prediction, the data stream includes readings from other key sensors, such as supply water temperature, return water temperature, flow rate, and pressure. These parameters are crucial information for understanding the operating status and load changes of the heating station, providing rich context for subsequent feature extraction. To construct long-term meso-flow characteristics, it is necessary to extract telemetry data over a relatively long period, for example, by setting the extraction of continuous telemetry data from the past 72 hours (3 days) or 168 hours (7 days).

[0018] Obtain the raw meteorological data stream. This stream is acquired from third-party commercial meteorological service APIs, such as Wind Weather, OpenWeatherMap, or data platforms published by meteorological bureaus. Compared to telemetry data, meteorological data is collected at a lower frequency, such as hourly, i.e., once per hour. This data stream contains key meteorological features that significantly affect heat load, such as outdoor temperature, wind speed, humidity, and solar radiation intensity. Specifically, the meteorological data needs to include two parts: historical meteorological data and future weather forecast data. Historical meteorological data is used for alignment with historical heat load data; for example, historical meteorological data for the past 72 hours prior to the forecast point is required. Future weather forecast data is used to construct the future context of the macro-flow; for example, weather forecast values ​​for the next 24 hours from the forecast point are required. In particular, these windows are pre-set fixed parameters based on an empirical understanding of the multi-timescale characteristics of heat load and forecasting needs, used to extract historical telemetry data and future weather forecast data for different time ranges.

[0019] In step S2, the original telemetry data stream from the heat station and the original meteorological data stream are preprocessed by splitting them to obtain micro-sequence input, meso-sequence input, and macro-context input for heat load prediction. Correspondingly, considering that changes in heat load are the result of superimposed signals at multiple time scales such as minutes, hours, and days, traditional methods often suffer from limited prediction accuracy due to a lack of refined processing of this multi-scale information. Therefore, this application splits the original data stream into micro-sequence input, meso-sequence input, and macro-context input for heat load prediction, laying the foundation for subsequent specialized feature extraction at different time granularities. This allows the model to comprehensively perceive the complex changes in heat load in short-term fluctuations, medium-term trends, and long-term periodicity. This splitting process avoids the redundancy and information loss caused by simple feature splicing in traditional methods and provides a prerequisite for the subsequent dynamic attention fusion mechanism to effectively identify and weigh the importance of signals at each scale, significantly improving the model's understanding of complex load changes and prediction accuracy.

[0020] Specifically, in an optional implementation, step S2, which involves splitting and preprocessing the original thermal station telemetry data stream and the original meteorological data stream to obtain the micro-sequence input, meso-sequence input, and macro-context input for heat load prediction, includes: S21, aligning and cleaning the original thermal station telemetry data stream and the original meteorological data stream to obtain the preprocessed original thermal station telemetry data stream and the preprocessed original meteorological data stream; S22, obtaining the timestamp of the prediction target as the prediction point; and S23, extracting the data before the prediction point from the preprocessed original thermal station telemetry data stream and the preprocessed original meteorological data stream. The original thermal station telemetry data and original meteorological data at each time step; 24, based on the prediction point before... The heat load prediction microsequence input is constructed from the raw thermal station telemetry data and raw meteorological data at each time step; S25, the data before the prediction point is extracted from the preprocessed raw thermal station telemetry data stream and the preprocessed raw meteorological data stream. S26, raw thermal station telemetry data and raw meteorological data at each time step; S26, for the time step before the predicted point The raw telemetry data of the heat station and the raw meteorological data at each time step are downsampled and aggregated to obtain the mesoscopic sequence input for the heat load prediction.

[0021] Specifically, in one optional implementation, the It lasts for 120 minutes. The timeframe is 72 hours. These values ​​are based on a deep understanding and empirical consideration of the multi-timescale characteristics of the heating load at the heating station. The 120-minute window aims to capture the microscopic, instantaneous changes in the heating load, as the heating load may fluctuate rapidly in a short period due to sudden temperature changes, sudden adjustments in user behavior, or real-time system scheduling. The 2-hour window is sufficient to cover these high-frequency dynamics, providing the model with a fine recent context, enabling it to respond to the most immediate load changes. The 72-hour window is used to capture mesoscale trends and periodic patterns, such as the obvious daily periodicity of the heating load (morning and evening peaks, nighttime troughs) and short-term trends spanning several days (such as temperature changes over several consecutive days or differences between weekdays and weekends). Three days of historical data provides a sufficient span to identify these recurring patterns and medium- to long-term trends, while avoiding the computational burden and irrelevant noise that may result from excessively long sequences.

[0022] Understandably, raw data often contains various quality issues, which, without preprocessing, will severely impact the accuracy and robustness of subsequent feature extraction and model prediction. Specifically, raw data streams may originate from different acquisition devices or data sources, leading to inconsistent timestamps and different sampling frequencies. For example, telemetry data may be minute-level, while meteorological data may be hour-level, making direct fusion and analysis difficult. Furthermore, sensor malfunctions, transmission interruptions, or environmental interference often result in missing or outlier values ​​in the data. These incomplete or erroneous data introduce noise, mislead model learning, and may even cause model collapse. Therefore, alignment and cleaning aim to eliminate inconsistencies, incompleteness, and errors in the data, ensuring that the data input to the model is high-quality, synchronized, and reliable, thus laying a solid foundation for accurate subsequent predictions.

[0023] Specifically, in an optional implementation, step S21, aligning and cleaning the original thermal station telemetry data stream and the original meteorological data stream to obtain preprocessed original thermal station telemetry data stream and preprocessed original meteorological data stream, includes: S211, aligning the original thermal station telemetry data stream and the original meteorological data stream with timestamps to obtain aligned original thermal station telemetry data stream and aligned original meteorological data stream; S212, interpolating the aligned original meteorological data stream to obtain interpolated original meteorological data stream; S213, handling missing values ​​and outliers in the aligned original thermal station telemetry data stream and the interpolated original meteorological data stream to obtain the preprocessed original thermal station telemetry data stream and the preprocessed original meteorological data stream.

[0024] Step S21 is achieved through the following process: First, proceed with S211. Since the original thermal station telemetry data stream is collected at a minute-level frequency (e.g., once per minute), while the original meteorological data stream is typically collected at an hour-level frequency (e.g., once per hour), there is a difference in their time granularity. To achieve data fusion, they need to be aligned to a common time reference. Specifically, using the minute-level timestamp as the primary reference, for each minute-level timestamp, the nearest hour-level timestamp data in the original meteorological data stream is searched and correlated. For example, if the thermal station telemetry data is recorded at times such as 08:01 and 08:02, while the meteorological data is only recorded at the top of the hour such as 08:00 and 09:00, then the telemetry data from 08:01 to 08:59 will be initially aligned with the meteorological data from 08:00, ensuring consistency in the time dimension of data from different sources and at different frequencies, providing a unified time view for subsequent processing.

[0025] Secondly, S212 is performed. Although timestamp alignment was performed in S211, due to the low-frequency nature of meteorological data, there may still be many gaps in the minute-level timestamps. For example, meteorological data is available at 08:00 and 09:00, but there are no direct meteorological observations between 08:01 and 08:59. To fill these gaps and provide more refined meteorological information, interpolation methods are used. Commonly used interpolation methods include linear interpolation and spline interpolation. For example, linear interpolation can be used to calculate the meteorological values ​​per minute from 08:01 to 08:59 based on the meteorological data at 08:00 and 09:00. For example, at 08:00, the outdoor temperature is recorded as 10.0℃; while at 09:00, the outdoor temperature is recorded as 12.0℃. Linear interpolation ensures that the data changes linearly and uniformly between two known time points. In practice, the total time span between the start time (08:00) and the end time (09:00) is first determined to be 60 minutes. Then, the temperature increment per minute is calculated, which is approximately 0.0333℃ / minute (12.0 - 10.0) / 60. Interpolation is then performed for any minute-level time point between 08:00 and 09:00. For 08:01, the temperature will be approximately 10.0℃ + (0.0333℃ / minute × 1 minute); and for 08:59, the temperature will be approximately 11.9647℃ + (0.0333℃ / minute × 59 minutes). Through this linear extrapolation, hourly gaps in the original meteorological data stream are filled with continuous minute-level data, ensuring a perfect temporal match with the thermal station telemetry data, thus obtaining the interpolated original meteorological data stream.

[0026] Finally, proceed to S213. During data acquisition and transmission, data gaps or anomalies are inevitable. For missing values, various strategies can be employed for filling. For example, for short-term gaps, the previous valid value can be used (forward filling), or the average of adjacent valid values ​​can be used. For long-term gaps, more complex interpolation may be necessary, combining historical data or domain knowledge. Outliers, i.e., data points that significantly deviate from the normal range, need to be identified and corrected. Identification methods can be based on statistical principles, such as Z-score, IQR, or preset physical thresholds. Correction methods include replacing outliers with the mean, median, or a reasonable value obtained through interpolation, or limiting them to a reasonable physical range. For example, if the water supply temperature suddenly shows an abnormal reading of 500℃, it can be corrected to the historical maximum or a reasonable value obtained through interpolation.

[0027] It is understandable that heat load forecasting is essentially an estimation of heating demand at a future point in time. Therefore, accurately defining this future point in time is the foundation for all subsequent data extraction, feature construction, and model training. Without this forecast point, the model will be unable to determine which time periods to extract from historical data, nor will it be able to clearly define the future time range for prediction, leading to chaotic input data and loss of timeliness and accuracy in the prediction results. This forecast point, as a time anchor, ensures that the historical data and future forecast data contained in subsequent micro- and meso-level sequence inputs, as well as macro-level context inputs, can be organized and aligned around this specific forecast point, enabling the model to learn the multi-timescale features most relevant to that forecast point.

[0028] In an optional implementation, step S22 is achieved through the following process: Specifically, the timestamp of the predicted target depends on the timeliness requirements of the actual application scenario. A common implementation is that the prediction point is dynamically determined by an external request or a preset scheduling plan. For example, in the daily operation of a heating station, it may be necessary to predict the heating load for the next hour every hour. In this case, the prediction point can be set as the current time plus a fixed prediction lead time. Specifically, if the current real-time time is 10:00:00 on February 16, 2025, and the heating load for the next hour needs to be predicted, then the prediction point will be determined as 11:00:00 on February 16, 2025. This prediction point is a timestamp accurate to the minute or second, such as 2025-02-16 11:00:00.

[0029] Another implementation method is that the forecast point can be manually entered by a human operator based on actual needs. For example, when a dispatcher needs to know the heating load under specific holidays or extreme weather conditions, they can manually specify a future date and time as the forecast point, such as 11:00:00 on February 16, 2024.

[0030] Correspondingly, instantaneous changes in heat load are most directly affected by the most recent operating status and meteorological conditions. Background technology indicates that heat load changes are the result of multi-scale signal superposition, and micro-scale (minute-level) signals are crucial for capturing rapid fluctuations. This application, by extracting high-frequency data from H-short (e.g., 120 minutes) before the prediction point, can capture the latest dynamics of heating station operating parameters and meteorological conditions, such as minor adjustments in water supply temperature and rapid rises and falls in outdoor temperature. This detailed information plays a decisive role in predicting short-term future heat load.

[0031] In an optional implementation, step S23 is implemented through the following process: First, the timestamp of the prediction point is determined. For example, if the prediction point is February 16, 2025, at 11:00:00. Then, according to a preset... The value (120 minutes) is used to calculate the starting timestamp of the data to be extracted. The starting timestamp equals the prediction point timestamp minus... Duration. For example, subtracting 120 minutes from 11:00:00 gives a start timestamp of 09:00:00 on February 16, 2025. Then, from the preprocessed raw telemetry data stream of the heating station, all data records with timestamps between the start timestamp and the prediction point timestamp are selected. Since the preprocessed telemetry data stream has a minute-level frequency, this will result in a telemetry data sequence containing 120 time steps, from 09:00 to 10:59. This sequence contains detailed changes in heating load, supply water temperature, return water temperature, flow rate, pressure, etc., over the past 120 minutes.

[0032] Simultaneously, data records with timestamps within the same time range were also selected from the preprocessed raw meteorological data stream. Since the preprocessed meteorological data stream has also been interpolated to a minute-level frequency, this will also result in a meteorological data sequence containing 120 time steps, including fine-grained changes in outdoor temperature, wind speed, humidity, and solar radiation intensity over the past 120 minutes. Through the above truncation operations, two segments of length are ultimately obtained. The raw telemetry data and raw meteorological data of the thermal station at each time step.

[0033] Understandably, as the background technology points out, changes in heat load are the result of superimposed signals at multiple time scales, among which minute-level micro-signals are crucial for understanding and predicting short-term load changes. Traditional prediction methods often struggle to effectively utilize this high-frequency, fine-grained data, leading to poor performance when dealing with rapidly changing loads. By constructing specially designed micro-sequence inputs, the latest dynamics of heating station operating parameters and meteorological conditions can be presented to the model in sequence form, such as minor adjustments in supply water temperature and rapid rises and falls in outdoor temperature within a short period. This detailed information directly reflects the current operating status of the heating system and the immediate impact of the external environment, providing high-quality data support for subsequent processing.

[0034] In an optional implementation, step S24 is achieved through the following process: For Within each time step of the time window, i.e., each minute, the corresponding features of the heating station telemetry data and meteorological data are concatenated. For example, in a certain minute, the heating station telemetry data includes the following features: heating load, supply water temperature, return water temperature, flow rate, and pressure. Simultaneously, the meteorological data includes the following features: outdoor temperature, wind speed, humidity, and solar radiation intensity. Therefore, in this minute, these feature values ​​are concatenated in a predetermined order to form a single feature vector. This concatenation operation is repeated for points prior to the prediction point. For each of the 120 minutes of data, feature concatenation is performed. Finally, these 120 concatenated feature vectors are arranged in chronological order to form a two-dimensional matrix, which is the micro-sequence input for heat load prediction. The dimension of this sequence will be [120, 9], where 120 represents the number of time steps and 9 represents the number of features at each time step.

[0035] As the background art points out, changes in heat load are not only affected by instantaneous factors, but are also significantly constrained by mesoscale patterns on an hourly and daily scale, such as daily peak and trough heat consumption, and differences between weekdays and weekends. Traditional forecasting methods often struggle to effectively capture these more macroscopic and periodic load patterns due to their limited receptive field. This application, by extracting longer historical data (H-long) prior to the forecast point, provides the model with a sufficient time span to learn these medium-scale patterns, such as load change trends over several consecutive days and the repetition of daily load curves. This provides a comprehensive historical context, enabling the model to more accurately understand and predict the overall trend of heat load, overcoming the shortcomings of existing models in capturing medium- and long-term load patterns.

[0036] In an optional implementation, step S25 is performed by the following process: First, the timestamp of the prediction point is determined. For example, if the prediction point is 11:00:00 on February 16, 2025.

[0037] Next, according to the preset The value, i.e., 72 hours, is used to calculate the starting timestamp of the data to be extracted. The starting timestamp equals the prediction point timestamp minus... Duration. For example, subtracting 72 hours from 11:00:00 gives a start timestamp of February 13, 2025, at 11:00:00. Then, from the preprocessed raw telemetry data stream, all data records with timestamps between the start timestamp and the prediction point timestamp are filtered out. Since the preprocessed telemetry data stream has a minute-level frequency, this will result in a... Multiplying by 60 time steps, this results in a telemetry data sequence of 72 hours * 60 minutes / hour = 4320 time steps. This sequence contains minute-level detailed changes in the operating parameters of the heating station over the past 72 hours.

[0038] Simultaneously, data records with timestamps within the same time range were also selected from the preprocessed raw meteorological data stream. Since the preprocessed meteorological data stream has also been interpolated to a minute-level frequency, this will also result in a meteorological data sequence containing 4320 time steps, encompassing minute-level fine changes in meteorological characteristics over the past 72 hours. Thus, two segments of length are ultimately obtained. Each time step consists of 72 hours of raw thermal station telemetry data and raw meteorological data.

[0039] Correspondingly, changes in heat load are not only affected by instantaneous factors, but also significantly constrained by mesoscale patterns at the hourly and daily levels, such as daily peak and trough heat usage, and differences between weekdays and weekends. S25 extracts minute-level data from the past 72 hours, which, while containing rich information, is too fine-grained for capturing hourly or daily trends, potentially introducing excessive noise and increasing the computational burden on subsequent model processing. Therefore, this application employs downsampling and aggregation at the hourly level, effectively smoothing short-term fluctuations and highlighting medium- to long-term trends. This allows the model to more clearly identify the daily and weekly periodicity of heat load, overcoming the shortcomings of existing models in capturing medium- to long-term load patterns and improving the model's training efficiency and generalization ability.

[0040] Specifically, in an optional implementation, step S26, before the predicted point The raw telemetry data and raw meteorological data of the heat station at each time step are downsampled and aggregated to obtain the mesoscopic sequence input for the heat load forecast, including: aggregated data hourly before the forecast point. The original telemetry data and original meteorological data of the heat station at each time step are downsampled to obtain the mesoscopic sequence input for the heat load prediction.

[0041] Step S26 is achieved through the following process: For each hour, all minute-level data within that hour are statistically aggregated. For example, for the hour preceding the prediction point (between 72 and 71 hours back from the prediction point), all raw telemetry data from the heating station within these 60 minutes, such as heating load, supply water temperature, return water temperature, flow rate, pressure, and raw meteorological data, such as outdoor temperature, wind speed, humidity, and solar radiation intensity, are aggregated separately. A commonly used aggregation method is to calculate the average value, as the average value can effectively smooth out instantaneous fluctuations in minute-level data and better reflect the overall trend within that hour. For example, the average heating load value for the hour is obtained by averaging the 60 minute-level heating load values ​​within that hour; similar average values ​​are also calculated for other characteristics such as outdoor temperature and wind speed.

[0042] Repeat the above aggregation operation for the predicted points. The feature aggregation is performed every hour of the 72-hour time step. For example, if the prediction point is 11:00:00 on February 16, 2025, the aggregation will start from 11:00:00 on February 13, 2025, and proceed hourly until 10:00:00 on February 16, 2025. Finally, these 72 aggregated hourly feature vectors are arranged in chronological order to form a two-dimensional matrix, which is the input of the mesoscopic sequence for the heat load prediction. The dimension of this sequence will be [72, 9], where 72 represents the number of hours and 9 represents the number of aggregated features for each hour.

[0043] Specifically, in an optional implementation, step S2, which involves splitting and preprocessing the original thermal station telemetry data stream and the original meteorological data stream to obtain the micro-sequence input, meso-sequence input, and macro-context input for heat load prediction, further includes: extracting time-related features based on the timestamp of the prediction target; extracting weather forecast values ​​for the next N hours from the original meteorological data stream; and combining the time-related features and the weather forecast values ​​for the next N hours to obtain the macro-context input for heat load prediction.

[0044] In an optional implementation, the macroscopic context input generation process for heat load prediction is implemented through the following steps: First, time-related features are extracted based on the timestamp of the prediction target. For example, if the prediction point is 11:00:00 on February 16, 2025, various features reflecting time periodicity and attributes can be extracted from this timestamp. These features include, but are not limited to: year, month, date, day of the week (e.g., Sunday can be encoded as 7), hour, whether it is a weekday (e.g., 1 indicates a weekday, 0 indicates a non-weekday), and whether it is a holiday (requires a pre-set holiday calendar for querying, e.g., 0 indicates a non-holiday). In addition, more abstract time features can be extracted, such as time periods of the day, such as morning, afternoon, and night, which can be individually encoded, or the stage of the heating season, such as the beginning, middle, and end of the heating season. These time-related features are combined into a time-related feature vector, for example, [2025, 2, 16, 7, 11, 1, 0, AM].

[0045] Secondly, weather forecast values ​​for the next N hours are extracted from the original meteorological data stream. The determination of N depends on the actual application's need for forecast lead time and the reliability of the weather forecast; for example, it can be set to 12 hours or 24 hours. In practice, hourly weather forecast data for the next N consecutive hours (e.g., 12 hours) starting from the target forecast timestamp is extracted from the original meteorological data stream. For each hour, key meteorological features are extracted, such as outdoor temperature, wind speed, humidity, and solar radiation intensity. For example, if the forecast point is 11:00:00 on February 16, 2025, then hourly weather forecast data for the 12 hours from 11:00:00 on February 16, 2025 to 22:00:00 on February 16, 2025 needs to be extracted. These data will form a sequence where each element represents an hourly meteorological feature vector, such as {[T1,W1,H1,S1],[T2,W2,H2,S2],...,[T12,W12,H12,S12]}, where T, W, H, and S represent temperature, wind speed, humidity, and solar radiation intensity, respectively.

[0046] Finally, the time-based features and the weather forecast values ​​for the next N hours are combined to obtain the macroscopic context input for heat load prediction. The time-based feature vector is concatenated with the sequence of weather forecast values ​​for the next N hours. The sequence of future weather forecast values ​​is flattened into a long vector, which is then concatenated with the time-based feature vector. The final macroscopic context input will be a single feature vector containing all macroscopic information, i.e., the macroscopic context input for heat load prediction.

[0047] In step S3, multi-timescale feature extraction is performed on the micro-scale sequence input, meso-scale sequence input, and macro-scale context input of heat load prediction to obtain micro-scale, meso-scale, and macro-scale embedding vectors for heat load prediction. It should be understood that, as clearly stated in the background art, heat load changes are a manifestation of the superposition of signals at multiple time scales, and existing models struggle to fully capture these characteristics. By designing a dedicated feature extractor for each time scale, the advantages of various neural networks in processing specific data patterns can be fully utilized. For example, CNN excels at capturing local temporal patterns, and LSTM excels at capturing long-distance dependencies. Transforming the raw data into embedding vectors not only reduces the dimensionality and redundancy of the data, but more importantly, these embedding vectors can abstract the most predictive potential features and patterns in the data at each scale. This effectively solves the redundancy and information loss problems caused by simple feature concatenation in traditional methods, laying the foundation for subsequent intelligent weighted fusion of information at different scales, thereby significantly improving the model's understanding and prediction accuracy of complex load changes.

[0048] Specifically, in one alternative implementation, Figure 3 This is a flowchart of step S3 in the intelligent prediction method for heating load of a heat station according to an embodiment of this application. Figure 3 As shown, step S3, which involves performing multi-timescale feature extraction on the heat load prediction micro-sequence input, the heat load prediction meso-sequence input, and the heat load prediction macro-context input to obtain heat load prediction micro-scale embedding vector, heat load prediction meso-scale embedding vector, and heat load prediction macro-scale embedding vector, includes: S31, inputting the heat load prediction micro-sequence into a micro-flow feature extractor based on a CNN network to obtain the heat load prediction micro-scale embedding vector; S32, inputting the heat load prediction meso-sequence into a meso-flow feature extractor based on an LSTM network to obtain the heat load prediction meso-scale embedding vector; and S33, performing macro-flow context encoding on the heat load prediction macro-context input to obtain the heat load prediction macro-scale embedding vector.

[0049] Step S3 is implemented through the following process: First, proceed to S31. The CNN network consists of multiple Conv1D layers, activation functions such as ReLU, and pooling layers such as global average pooling. For example, the network may contain: an input layer that receives a sequence of heat load prediction microscale [120,9]; followed by multiple Conv1D layers, each using convolutional kernels of different sizes, for example, kernel sizes of 3, 5, and 7, to capture local patterns of different lengths in the microscale sequence, such as load change trends or weather condition fluctuations over consecutive 3, 5, or 7 minutes. Each convolutional layer is followed by a ReLU activation function, introducing non-linearity. Subsequently, a global average pooling layer can be used to reduce the dimensionality of the feature maps extracted by each convolutional kernel to a single feature vector, thereby summarizing the local features of the entire microscale sequence. Finally, a fully connected layer maps these summarized features into a fixed-dimensional heat load prediction microscale embedding vector.

[0050] Next, S32 is performed. The LSTM network consists of one or more LSTM layers that can effectively handle long-term dependencies in sequence data. For example, the network may contain: an input layer that receives the mesoscale sequence [72, 9] of heat load prediction as input; followed by one or more LSTM layers, each containing multiple LSTM units that selectively memorize or forget information through gating mechanisms (input gate, forget gate, output gate), thereby capturing hourly trends and periodic patterns in heat load and meteorological data over the past 72 hours, such as the evolution of daily load peaks and troughs. After the LSTM layers process the sequence, only the hidden state of the last time step is taken as the representation of the entire sequence, or the hidden states of all time steps are pooled. Finally, a fully connected layer maps the features extracted by the LSTM to a fixed-dimensional mesoscale embedding vector for heat load prediction.

[0051] Finally, S33 is performed. The input is fed into the macroscopic flow context encoding module, which is a simple fully connected neural network. For example, this module might contain: an input layer that receives the macroscopic context input for heat load prediction; followed by one or more fully connected layers, each followed by an activation function such as ReLU, to learn the non-linear relationships between these macroscopic features and transform them into a more abstract and expressive representation. Finally, an output fully connected layer maps the processed features into a fixed-dimensional macroscopic scale embedding vector for heat load prediction. Specifically, the weights and bias parameters in the above module network are obtained during model training.

[0052] In step S4, the micro-scale embedding vector, meso-scale embedding vector, and macro-scale embedding vector of heat load prediction are input into a dynamic attention fusion gate to obtain a multi-scale fusion embedding vector for heat load prediction. Correspondingly, the accuracy of heat load prediction highly depends on the comprehensive utilization of information at different time scales. However, the importance of information at each scale is not fixed but dynamically changes with the external environment and operating status. For example, when the load fluctuates drastically, micro-scale information may be more critical; while when the load is stable or when predicting long-term trends, meso- and macro-scale information may dominate. Background technology points out that existing methods lack a dynamic weight adjustment mechanism, resulting in insufficient prediction accuracy and robustness. Based on this, this application introduces a dynamic attention fusion gate, enabling the model to automatically learn and assign different weights to each scale embedding vector according to the context of the current input data, thereby dynamically emphasizing the most relevant feature information. This effectively solves the limitations of fixed weights or simple concatenation in traditional methods, significantly improving the model's adaptability to complex load changes and prediction accuracy.

[0053] Specifically, in one alternative implementation, Figure 4 This is a flowchart of step S4 in the intelligent prediction method for heating load of a heating station according to an embodiment of this application. Figure 4As shown, step S4, which involves inputting the microscale embedding vector, mesoscale embedding vector, and macroscale embedding vector of heat load prediction into a dynamic attention fusion gate to obtain a multiscale fusion embedding vector of heat load prediction, includes: S41, concatenating the microscale embedding vector, mesoscale embedding vector, and macroscale embedding vector of heat load prediction into a multiscale concatenated embedding vector of heat load prediction; S42, inputting the multiscale concatenated embedding vector of heat load prediction into a small feedforward neural network to obtain microscale weight values, mesoscale weight values, and macroscale weight values; and S43, using the microscale weight values, mesoscale weight values, and macroscale weight values ​​to perform weighted aggregation on the microscale embedding vector, mesoscale embedding vector, and macroscale embedding vector of heat load prediction to obtain the multiscale fusion embedding vector of heat load prediction.

[0054] Step S4 is achieved through the following process: First, proceed with S41. The three fixed-dimensional vectors are simply concatenated along their feature dimensions. For example, by connecting the micro-scale embedding vector, meso-scale embedding vector, and macro-scale embedding vector of heat load prediction along their dimensions, the resulting multi-scale concatenated embedding vector of heat load prediction will be a concatenation of these three vectors, integrating abstract features from different time scales into a unified vector space.

[0055] Next, proceed to S42. The architecture of a small feedforward neural network comprises one or more fully connected layers and activation functions. For example, the network can be designed as follows: an input layer that receives a high-dimensional heat load prediction multi-scale concatenated embedding vector; one or more hidden fully connected layers, for example, the first hidden layer can map the high dimension to an intermediate dimension and apply the Modified Linear Unit (ReLU) activation function to introduce non-linearity; and an output fully connected layer with three neurons, corresponding to weights at the micro, meso, and macro scales. To ensure that these weights are positive and sum to 1, a Softmax activation function is applied to this output layer. The Softmax function normalizes the output values ​​to a probability distribution between 0 and 1, thus obtaining the micro-scale weights, meso-scale weights, and macro-scale weights.

[0056] Specifically, the weight and bias parameters of the feedforward neural network are learned end-to-end throughout the entire model training process.

[0057] Finally, proceed to S43. Weighted aggregation involves element-wise multiplication of each scale embedding vector with its corresponding scale weight, followed by summing these weighted vectors:

[0058] ;in, It is a microscale embedding vector for heat load prediction. The mesoscale embedding vector for heat load prediction is: It is the macroscopic embedding vector for heat load prediction. and These are the microscale weight value, the mesoscale weight value, and the macroscale weight value, respectively. It is added based on the position point. It is a multi-scale fusion embedding vector for heat load prediction. This fusion vector integrates information from different time scales and weights them according to their dynamic importance, thus forming a more comprehensive and representative feature representation, providing high-quality input for subsequent prediction heads.

[0059] In step S5, the multi-scale fusion embedding vector for heat load prediction is input into the prediction head based on the regressor to obtain the final heat load prediction value. Accordingly, heat load prediction is essentially a regression problem, requiring a continuous numerical output. As mentioned in the background, existing prediction methods lack accuracy, while this scheme constructs a highly condensed and semantically rich fusion embedding vector through multi-scale feature extraction and dynamic attention fusion. This step is the endpoint of the entire prediction process; it maps this high-dimensional, abstract feature representation to the actual heat load numerical space, thereby achieving accurate quantitative prediction of future heating demand and providing direct decision-making basis for the operation and scheduling of heating stations, ensuring the practicality and effectiveness of the entire method.

[0060] It is worth noting that although the dynamic attention fusion gate intelligently weights the embedding vectors at each scale, the weighted feature distances between feature vectors may still differ significantly after simple weighting due to the different encoding methods and information granularities of micro, meso, and macro scale features. This can lead to potential inconsistencies or information conflicts within the fusion vector. If this inherent heterogeneity is not addressed, it may cause deviations in subsequent regression predictions or even regression aggregation failure. Therefore, before decoding, the multi-scale fusion embedding vectors for heat load prediction obtained through weighted aggregation need to be optimized to further refine and correct them, eliminating or mitigating the negative impact of these heterogeneous encoding differences. This improves the expressive power and robustness of the fusion vectors, ensuring the accuracy and stability of the final heat load prediction.

[0061] Specifically, in an optional implementation, the heat load prediction multi-scale fusion embedding vector is input into the prediction head based on a regressor to obtain the final heat load prediction value, including: first, based on the micro-scale weight value, the meso-scale weight value, and the macro-scale weight value, the heat load prediction multi-scale fusion embedding vector is subjected to eigenvalue-based weight distance calculation to obtain the heat load prediction micro-weight distance, the heat load prediction meso-weight distance, and the heat load prediction macro-weight distance, that is: ;in, It is the third among the micro-scale weight value, meso-scale weight value, and macro-scale weight value. Each scale weight value, To predict the length of the multi-scale fused embedding vector for heat load, These are the feature values ​​in the multi-scale fusion embedding vector for heat load prediction. It is the third of the three weighted distances for heat load forecasting: micro-weighted distance, meso-weighted distance, and macro-weighted distance. Each scale weight value has a weight distance. It should be understood that, as described above, each scale embedding vector is assigned dynamic weights and weighted. However, this aggregation may not completely eliminate the inherent heterogeneity of features at different scales. By calculating the weight distance between each scale weight value and the corresponding feature value in the multi-scale fusion embedding vector of heat load prediction, the difference or deviation of each scale's contribution to the fusion vector can be accurately measured. Each weight distance is a numerical value that quantifies the difference between the weighted contribution of each scale and the features of the fusion vector, serving as a measure of the scale and heterogeneous encoding integration difference of the feature values. This provides a clear and operational difference primitive for subsequent optimization.

[0062] Next, the micro-weight distance, meso-weight distance, and macro-weight distance of the heat load prediction are used as integrated difference primitive modules to perform decoding target prediction primitive integration on the multi-scale fusion embedding vector of the heat load prediction to obtain the multi-scale fusion embedding integrated vector of the heat load prediction, that is: ;in, It is added based on the position point. It is a multi-scale fusion embedding vector for heat load prediction. It is a multi-scale fusion embedding integrated vector for heat load prediction. That is, by obtaining the integrated difference primitives, i.e. Each primitive module can be mapped to a decoding target submodule, thus obtaining the differential prediction regression driving vector under the decoding target through direct integration. This means... This is used to weight or scale the original multi-scale fusion embedding vector of heat load prediction to highlight or adjust information that may be obscured due to heterogeneity. The resulting multi-scale fusion embedding vector of heat load prediction is a difference-driven fusion vector that can more clearly reflect the heterogeneity effects that may exist in the aggregation process, providing another perspective for subsequent interactive corrections.

[0063] Then, a cosine-like correlation interaction is performed on the heat load prediction multi-scale fusion embedding vector and the heat load prediction multi-scale fusion embedding integration vector to obtain the heat load prediction modeling interaction matrix, namely: ;in, It is a vector transpose operation. It calculates the L2 norm of a vector. It is vector multiplication. It is the interaction matrix for heat load prediction modeling. These are column vectors. It's understandable that the complexity of heat load requires the model to consider not only the features themselves but also the interactions between them. Using a cosine-like correlation interaction can measure the directional similarity between these two vectors and transform it into an interaction matrix. This matrix can capture... and The interaction matrix for heat load prediction modeling encodes the complex relationship between the original fused features and the differential driving features by analyzing the strength of the correlations and the direction of their influence across various dimensions. This matrix will serve as a key element in subsequent correction operations, guiding how to adjust the original fused vector.

[0064] Then, the heat load prediction multi-scale fusion embedding vector is corrected using the heat load prediction modeling interaction matrix to obtain the corrected heat load prediction multi-scale fusion embedding vector, i.e.: ;in, This is a modified multi-scale fusion embedding vector for heat load prediction. Thus, matrix multiplication is used to... pass Transform to thus and The dynamic interaction between them is integrated into This avoids aggregation regression failure caused by heterogeneity; that is, through this interactive correction, the negative impact of feature heterogeneity at different scales is eliminated or mitigated. This results in a modified heat load prediction multi-scale fusion embedding vector that not only includes dynamic fusion of multi-scale information but also explicitly models and corrects heterogeneous coding differences, thus providing higher-quality input for the final heat load prediction and significantly improving the model's predictive performance and stability.

[0065] Finally, the corrected heat load prediction multi-scale fusion embedding vector is input into the prediction head based on the regressor to obtain the final heat load prediction value. In other words, after the fine optimization of the preceding steps, this corrected vector has highly condensed and corrected heat load-related information from different time scales (micro, meso, and macro), effectively solving the challenges of heterogeneous encoding and aggregation decoding. At this point, the vector represents the most comprehensive and accurate abstract feature representation of the future heat load. Therefore, inputting it into a prediction head specifically designed for regression tasks can efficiently map these abstract features to a specific, continuous heat load numerical space, thereby directly outputting the final, high-precision heat load prediction value, providing a reliable decision-making basis for the actual operation and scheduling of the heating station.

[0066] The specific implementation process of inputting the modified heat load prediction multi-scale fusion embedding vector into a regressor-based prediction head to obtain the final heat load prediction value is as follows: The modified heat load prediction multi-scale fusion embedding vector is input into a regressor-based prediction head. This prediction head is a small feedforward neural network used to map high-dimensional features to a single continuous output value. Its architecture includes: Input layer: receiving the heat load prediction multi-scale fusion embedding vector. Hidden layer, containing one or more fully connected layers. For example, the first hidden layer can map the input dimension from 128 to an intermediate dimension (such as 64-dimensional) and apply non-linear activation functions such as the Modified Linear Unit (ReLU) to learn the complex non-linear relationship between the fused features and the heat load. If there are multiple hidden layers, they will continue to perform deeper abstraction and transformation of the features. Output layer contains a single neuron for outputting the final heat load prediction value. Since the heat load is a continuous value.

[0067] During the encoding process, the modified heat load prediction multi-scale fusion embedding vector first undergoes a linear transformation through the first fully connected layer of the prediction head, i.e., multiplied by the weight matrix of this layer and supplemented with a bias vector, and then non-linearity is introduced through an activation function. This process is repeated in subsequent hidden layers, with each layer taking the output of the previous layer as input to further extract and transform features. Finally, after processing by all hidden layers, the last fully connected layer maps these highly abstract features into a single numerical value, i.e., the final heat load prediction value. Notably, the weights and biases of all fully connected layers in this prediction head are learnable parameters obtained through training.

[0068] In summary, the intelligent prediction method for heating load of heat stations based on the embodiments of this application is explained. First, it performs refined preprocessing of the original telemetry and meteorological data from the heat station, generating micro-sequence input, meso-sequence input, and macro-context input for heat load prediction. This captures information at different time granularities at the data level, overcoming the shortcomings of traditional methods in processing multi-scale data. Subsequently, feature extraction is performed on the inputs at each scale to obtain their respective embedding vectors. The key lies in introducing a dynamic attention fusion gate. This module can intelligently learn and dynamically allocate the weights of the embedding vectors at each scale, effectively resolving conflicts and priority issues between signals at different time scales, avoiding the limitations of hard coding in traditional methods, and expanding the model's receptive field. Finally, the fused multi-scale information is input into a regressor for prediction, significantly improving the accuracy and robustness of the prediction, overcoming the shortcomings of existing schemes in effectively handling multi-scale dynamic changes.

[0069] Figure 5 This is a block diagram of a heating station intelligent load prediction system according to an embodiment of this application. Figure 5 As shown, the intelligent heating load prediction system 100 for heating stations according to an embodiment of this application includes: a heating station weather data acquisition module 110, used to acquire raw heating station telemetry data streams and raw meteorological data streams; a data splitting and preprocessing module 120, used to split and preprocess the raw heating station telemetry data streams and the raw meteorological data streams to obtain micro-sequence input, meso-sequence input, and macro-context input for heating load prediction; and a multi-scale heating load prediction encoding module 130, used to encode the micro-sequence input, meso-sequence input, and macro-context input for heating load prediction. The heat load prediction module 140 performs multi-timescale feature extraction on the context input to obtain micro-scale embedding vectors, meso-scale embedding vectors, and macro-scale embedding vectors. The heat load prediction multi-scale fusion module 150 is used to input the heat load prediction micro-scale embedding vectors, meso-scale embedding vectors, and macro-scale embedding vectors into a dynamic attention fusion gate to obtain a heat load prediction multi-scale fused embedding vector. The heat load prediction module 150 is used to input the heat load prediction multi-scale fused embedding vector into a regression-based prediction head to obtain the final heat load prediction value.

[0070] Here, those skilled in the art will understand that the specific operations of each step in the above-mentioned intelligent prediction system for heating load of heating stations have been referenced above. Figures 1 to 4 The description of the intelligent prediction method for heating load of heat stations is detailed here, and therefore, its repeated description will be omitted.

Claims

1. A heat supply load intelligent prediction method for a heat station, characterized in that, The method comprises the following steps: obtaining an original heat station telemetry data stream and an original meteorological data stream; performing stream preprocessing on the original heat station telemetry data stream and the original meteorological data stream to obtain a heat load prediction microscopic sequence input, a heat load prediction mesoscopic sequence input and a heat load prediction macroscopic context input; performing multi-time scale feature extraction on the heat load prediction microscopic sequence input, the heat load prediction mesoscopic sequence input and the heat load prediction macroscopic context input to obtain a heat load prediction microscopic scale embedding vector, a heat load prediction mesoscopic scale embedding vector and a heat load prediction macroscopic scale embedding vector; inputting the heat load prediction microscopic scale embedding vector, the heat load prediction mesoscopic scale embedding vector and the heat load prediction macroscopic scale embedding vector into a dynamic attention fusion gate to obtain a heat load prediction multi-scale fusion embedding vector, comprising: splicing the heat load prediction microscopic scale embedding vector, the heat load prediction mesoscopic scale embedding vector and the heat load prediction macroscopic scale embedding vector into a heat load prediction multi-scale spliced embedding vector; inputting the heat load prediction multi-scale spliced embedding vector into a small feedforward neural network to obtain a microscopic scale weight value, a mesoscopic scale weight value and a macroscopic scale weight value; weighting and aggregating the heat load prediction microscopic scale embedding vector, the heat load prediction mesoscopic scale embedding vector and the heat load prediction macroscopic scale embedding vector by using the microscopic scale weight value, the mesoscopic scale weight value and the macroscopic scale weight value to obtain the heat load prediction multi-scale fusion embedding vector; inputting the heat load prediction multi-scale fusion embedding vector into a regressor-based prediction head to obtain a final heat load prediction value.

2. The heat supply station heat load intelligent prediction method according to claim 1, characterized in that, The original heat station telemetry data stream and the original weather data stream are preprocessed to obtain a heat load prediction micro sequence input, a heat load prediction meso sequence input and a heat load prediction macro context input, including: aligning and cleaning the original heat station telemetry data stream and the original weather data stream to obtain preprocessed original heat station telemetry data stream and preprocessed original weather data stream; obtaining the timestamp of the prediction target as a prediction point; intercepting the original heat station telemetry data and the original weather data of the prediction point before time steps from the preprocessed original heat station telemetry data stream and the preprocessed original weather data stream; constructing the heat load prediction micro sequence input based on the original heat station telemetry data and the original weather data of the prediction point before time steps; intercepting the original heat station telemetry data and the original weather data of the prediction point before time steps from the preprocessed original heat station telemetry data stream and the preprocessed original weather data stream; and performing down-sampling aggregation processing on the original heat station telemetry data and the original weather data of the prediction point before time steps to obtain the heat load prediction meso sequence input.

3. The heat supply station heat load intelligent prediction method according to claim 2, characterized in that, aligning and cleaning the original heat station telemetry data stream and the original meteorological data stream to obtain a preprocessed original heat station telemetry data stream and a preprocessed original meteorological data stream, comprising: performing timestamp alignment on the original heat station telemetry data stream and the original meteorological data stream to obtain an aligned original heat station telemetry data stream and an aligned original meteorological data stream; performing interpolation processing on the aligned original meteorological data stream to obtain an interpolated original meteorological data stream; performing missing value processing and outlier processing on the aligned original heat station telemetry data stream and the interpolated original meteorological data stream to obtain the preprocessed original heat station telemetry data stream and the preprocessed original meteorological data stream. 4.The heat-supply load intelligent prediction method of claim 2, wherein, The is 120 minutes, is 72 hours; wherein the original heat station telemetry data and original weather data of the preceding time steps for the prediction point are down-sampled and aggregated to obtain the macro-sequence input for the heat load prediction, comprising: down-sampling the original heat station telemetry data and original weather data of the preceding time steps for the prediction point to obtain the macro-sequence input for the heat load prediction. 5.The heat-supply load intelligent prediction method of a heat station according to claim 2, characterized in that, performing stream preprocessing on the original heat station telemetry data stream and the original meteorological data stream to obtain a heat load prediction microscopic sequence input, a heat load prediction mesoscopic sequence input and a heat load prediction macroscopic context input, further comprising: extracting a time class feature based on the timestamp of the prediction target; extracting a weather forecast value of N hours in the future from the original meteorological data stream; combining the time class feature and the weather forecast value of N hours in the future to obtain the heat load prediction macroscopic context input. 6.The heat-supply load intelligent prediction method of a heat station according to claim 1, characterized in that, The heat load prediction micro sequence input, the heat load prediction meso sequence input and the heat load prediction macro context input are subjected to multi-time scale feature extraction to obtain a heat load prediction micro scale embedding vector, a heat load prediction meso scale embedding vector and a heat load prediction macro scale embedding vector, including: the heat load prediction micro sequence input is subjected to micro flow feature extraction based on a CNN network to obtain the heat load prediction micro scale embedding vector; the heat load prediction meso sequence input is subjected to meso flow feature extraction based on an LSTM network to obtain the heat load prediction meso scale embedding vector; the heat load prediction macro context input is subjected to macro flow context coding to obtain the heat load prediction macro scale embedding vector.

7. The heat supply station heat load intelligent prediction method according to claim 6, characterized in that, The heat load prediction multi-scale fusion embedding vector is input into a prediction head based on a regressor to obtain a final heat load prediction value, including: based on the micro scale weight value, the meso scale weight value and the macro scale weight value, the heat load prediction multi-scale fusion embedding vector is subjected to feature value-based weight distance calculation to obtain a heat load prediction micro weight distance, a heat load prediction meso weight distance and a heat load prediction macro weight distance; the heat load prediction micro weight distance, the heat load prediction meso weight distance and the heat load prediction macro weight distance are taken as an integrated difference element module to integrate decoding target prediction elements of the heat load prediction multi-scale fusion embedding vector to obtain a heat load prediction multi-scale fusion embedding integrated vector; the heat load prediction multi-scale fusion embedding vector and the heat load prediction multi-scale fusion embedding integrated vector are subjected to class cosine correlation interaction to obtain a heat load prediction modeling interaction matrix; the heat load prediction multi-scale fusion embedding vector is corrected by the heat load prediction modeling interaction matrix to obtain a corrected heat load prediction multi-scale fusion embedding vector; the corrected heat load prediction multi-scale fusion embedding vector is input into the prediction head based on the regressor to obtain the final heat load prediction value.

8. A heat supply station heat supply load intelligent prediction system, characterized in that, Including: A heat station weather data acquisition module is configured to acquire an original heat station telemetry data stream and an original meteorological data stream; A data shunting preprocessing module is configured to perform shunting preprocessing on the original heat station telemetry data stream and the original meteorological data stream to obtain a heat load prediction micro sequence input, a heat load prediction meso sequence input and a heat load prediction macro context input; A heat load multi-scale prediction coding module is configured to perform multi-time scale feature extraction on the heat load prediction micro sequence input, the heat load prediction meso sequence input and the heat load prediction macro context input to obtain a heat load prediction micro scale embedding vector, a heat load prediction meso scale embedding vector and a heat load prediction macro scale embedding vector; The thermal load prediction multi-scale fusion module is configured to input the thermal load prediction micro-scale embedding vector, the thermal load prediction meso-scale embedding vector and the thermal load prediction macro-scale embedding vector into a dynamic attention fusion gate to obtain a thermal load prediction multi-scale fusion embedding vector, and includes: splicing the thermal load prediction micro-scale embedding vector, the thermal load prediction meso-scale embedding vector and the thermal load prediction macro-scale embedding vector into a thermal load prediction multi-scale spliced embedding vector; inputting the thermal load prediction multi-scale spliced embedding vector into a small feedforward neural network to obtain a micro-scale weight value, a meso-scale weight value and a macro-scale weight value; weighting and aggregating the thermal load prediction micro-scale embedding vector, the thermal load prediction meso-scale embedding vector and the thermal load prediction macro-scale embedding vector according to the micro-scale weight value, the meso-scale weight value and the macro-scale weight value to obtain the thermal load prediction multi-scale fusion embedding vector; The thermal load prediction module is configured to input the thermal load prediction multi-scale fusion embedding vector into a regressor-based prediction head to obtain a final thermal load prediction value.

Images