Big data-based intelligent heating load prediction method and system

Abstract

This invention discloses a smart heating load forecasting method and system based on big data, belonging to the field of smart heating load forecasting technology. It includes: acquiring historical multi-source heterogeneous time-series data of a target area, performing time alignment and outlier removal to generate standardized time-series data; mapping the standardized time-series data to spatial grid cells according to the spatial topology of the heating area to construct a spatiotemporal correlation data matrix; extracting training sample segments through a sliding time window; extracting the temporal statistical features and spatial neighborhood features of each segment, and classifying the segments to determine load fluctuation patterns; using the fluctuation patterns as labels and iteratively training the initial prediction model to obtain a basic load forecasting model; acquiring real-time weather forecasts and heating operation data, processing them in the same way to generate real-time standardized data, and outputting the predicted heat load value for a future preset time point through the basic load forecasting model. The system executes the above method steps.

Description

Technical Field

[0001] This invention relates to the field of intelligent heating load prediction technology, specifically to an intelligent heating load prediction method and system based on big data. Background Technology

[0002] In the operation and management of centralized heating systems, accurate prediction of heat load is crucial for ensuring heating quality and reducing energy consumption. Existing heat load prediction technologies largely rely on time-series data from individual heating stations or pipe networks, using statistical models or conventional neural networks to extrapolate future heat loads. These approaches typically treat data from each monitoring point as independent time series, neglecting the physical spatial coupling of the heating pipe network and the spatiotemporal interactions of heat transport between different stations. This results in prediction models failing to effectively capture the dynamic changes in the overall regional heat load. The shortcomings of existing technologies lie in the fact that heat load fluctuations are driven not only by outdoor meteorological conditions but also by the spatial constraints of the heating pipe network's topological connectivity and the operating status of adjacent heating stations. Conventional methods only extract features from the time-series data of local monitoring points, lacking explicit modeling of spatial neighborhood information and spatiotemporal correlations, leading to significantly increased prediction bias when the spatial distribution of heat load changes rapidly. On the other hand, existing data-driven models often directly learn from the original time series segments end-to-end, failing to distinguish the differences in thermal response mechanisms under different load fluctuation modes. This makes the models insufficiently adaptable to non-stationary operating conditions such as sudden load increases and decreases, and prone to lag or overshoot when switching operating modes.

[0003] The problem addressed by this invention is how to incorporate the spatial dependency information of pipeline nodes implied by the spatial topology of the heating area into the prediction model, and how to classify the load change process using time-series statistical features and spatial neighborhood features, so that the model can establish differentiated mapping relationships for different fluctuation patterns, thereby obtaining more reliable heat load prediction results under complex meteorological changes and variable operating conditions. Summary of the Invention

[0004] This invention provides a smart heating load forecasting method and system based on big data. The purpose is to identify load fluctuation patterns by constructing a spatiotemporal correlation data matrix and extracting time-series statistical features and spatial neighborhood features, thereby improving the accuracy of heating load forecasting under spatial distribution and dynamic change conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a smart heating load prediction method based on big data, comprising:

[0006] Acquire multi-source heterogeneous time-series data of the target region within a historical time period, and perform time alignment and outlier removal on the multi-source heterogeneous time-series data to generate standardized time-series data;

[0007] Based on the preset spatial topology of the heating area, the standardized time-series data is mapped to the corresponding spatial grid cells to construct a spatiotemporal correlation data matrix;

[0008] Multiple training sample segments are generated by segmenting the spatiotemporal correlation data matrix using a sliding time window.

[0009] Extract the temporal statistical features and spatial neighborhood features of each training sample segment, classify the training sample segments according to the temporal statistical features and spatial neighborhood features, and determine the load fluctuation pattern corresponding to each training sample segment;

[0010] The load fluctuation pattern is used as a sample label, and iterative training is performed in combination with a preset initial prediction model to obtain a basic load prediction model.

[0011] Acquire real-time weather forecast data and real-time heating system operation data for the target area, and generate real-time standardized data according to the time alignment and outlier removal steps.

[0012] Based on the real-time standardized data, the load prediction model is used to perform prediction calculations and output the predicted heat load value of the target area at a future preset time point.

[0013] Preferably, the multi-source heterogeneous time-series data includes at least outdoor temperature, outdoor humidity, wind speed, solar radiation intensity, return water temperature of each heating station, supply water pressure of each heating station, and flow data of the heating network, so as to comprehensively capture external meteorological factors and internal operating status information that affect the heating load.

[0014] As a technical solution of the present invention, the time alignment and outlier removal specifically includes: using a preset unified sampling frequency as a benchmark, performing linear interpolation on data points with different timestamps to align all data sources to the same time axis; for each dimension of time-series data after alignment, calculating its local moving average and local moving standard deviation; when the value of a data point exceeds the sum of the corresponding local moving average and the local moving standard deviation under a preset multiple threshold, the data point is marked as an outlier to be checked and removed, and the average of the valid data points before and after the removal position is used to fill in the gaps to generate the standardized time-series data. This method effectively suppresses abnormal jumps caused by sensor drift, communication jitter, etc., ensuring the stability and reliability of the basic data.

[0015] Furthermore, the process of constructing the spatiotemporal correlation data matrix includes: parsing the spatial topology of the heating area, obtaining the geographic coordinates of heating pipeline nodes and heating stations, dividing the target area into grids, and determining the boundaries of spatial grid units; allocating standardized time-series data to corresponding grid units according to geographic location, arranging the data within each grid unit according to the time dimension to form a time-series vector, and combining the time-series vectors of all grid units according to spatial topological adjacency relationships to form the spatiotemporal correlation data matrix. This organically integrates the temporal dynamic characteristics and spatial distribution characteristics of the heating system, providing a structured data representation for exploring the spatiotemporal correlation patterns of heating load.

[0016] When using a sliding time window for extraction, a window length and a sliding step size are set, with the length being greater than the step size. The window is gradually moved starting from the time start point of the spatiotemporal correlation data matrix, extracting a data submatrix within the current time window as a training sample segment each time, until the entire time range is covered. The sample segments generated by this method can fully cover different heating conditions, such as stable and fluctuating periods.

[0017] When determining load fluctuation patterns, the peak, trough, rate of change, and dispersion (measured by variance or standard deviation) of the time series for each training sample segment are calculated and combined to form time-series statistical features. Simultaneously, the sum of the absolute values ​​of the differences between the central grid cell and all adjacent grid cells at the same time is calculated as a spatial neighborhood feature. The time-series statistical features and spatial neighborhood features are combined into a feature vector. The Euclidean distance between this feature vector and the center vectors of multiple preset load fluctuation patterns is calculated, and the pattern with the smallest Euclidean distance is determined as the load fluctuation pattern for that segment. Preferably, the multiple load fluctuation patterns include a stable pattern, a gradual increase pattern, a gradual decrease pattern, a sudden increase pattern, and a sudden decrease pattern. By incorporating spatial neighborhood differences, the transmission and coupling effects of heat load between adjacent areas can be more sensitively reflected, improving the ability to distinguish complex fluctuation patterns.

[0018] During the model training phase, the training sample segments are used as input, and their corresponding load fluctuation patterns are used as supervision information. These are fed into a pre-defined initial prediction model. The cross-entropy loss value between the predicted fluctuation pattern output by the model and the supervision information is calculated. Gradient descent is used for backpropagation to update the model parameters. This process is iterated repeatedly until the change in the loss value over multiple consecutive iterations is less than a pre-defined convergence threshold, thus obtaining the basic load prediction model. Training is guided by load fluctuation pattern classification, enabling the model to understand the inherent patterns of different trends during the learning process and enhancing its prediction robustness under rapid changes in heat load.

[0019] In real-time forecasting, the system receives meteorological forecast data for the target area at multiple points before a preset future time, as well as real-time operational data from various monitoring points of the heating system. It then sequentially performs time alignment, identification and removal of local moving average anomalies, and filling based on the mean of preceding and following valid data to generate real-time standardized data. This standardized data is mapped to spatial grid cells according to the spatial topology of the heating area, generating a real-time spatiotemporal correlation data matrix. A sliding time window is used to extract the forecast input segment, driving the load forecasting basic model to perform forward propagation calculations. The system outputs the predicted heat load values ​​for each spatial grid cell corresponding to the target area at the preset future time. By integrating real-time meteorological forecasts and system status, high spatiotemporal resolution dynamic heat load forecasting is achieved, providing refined data support for heating scheduling.

[0020] As a further improvement of this invention, after generating the predicted heat load value, a dynamic model correction step is also performed: recording the actual heat load value of the target area at a future time point, calculating the prediction deviation value between the actual heat load value and the predicted heat load value; when the absolute value of the prediction deviation value exceeds a preset deviation tolerance threshold multiple times consecutively, an online learning mechanism is triggered, adding the latest data containing the actual heat load value to the historical dataset, and re-performing standardization and model training operations to update the parameters of the load prediction base model online. Through online closed-loop correction, the model can continuously adapt to time-varying factors such as climate change, equipment aging, and evolution of building thermal characteristics, maintaining long-term high-precision prediction capabilities.

[0021] This invention also provides a smart heating load forecasting system based on big data, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned smart heating load forecasting method based on big data. This system can integrate multi-source heterogeneous heating network data, and while ensuring data quality, it accurately outputs future heat load forecasts using spatiotemporal correlation modeling and fluctuation pattern recognition. It also has adaptive update capabilities and can be widely applied to the scheduling optimization and energy-saving control of urban centralized heating systems.

[0022] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0023] By using a pre-defined spatial topology of the heating area, standardized time-series data is mapped to corresponding spatial grid cells to construct a spatiotemporal correlated data matrix. This processing method starts with the geographical coordinates of heating pipeline nodes and heating stations, divides the target area into grids, and assigns each data point to a spatial grid cell based on its geographical location. Then, the time-series vectors of each grid cell are generated by arranging them according to the time dimension. Finally, the data is combined according to spatial topological adjacency relationships to form a spatiotemporal correlated data matrix. This approach deeply integrates the physical connections of the heating system with the data organization, ensuring that the matrix not only contains the temporal evolution information of each cell but also fully preserves the spatial adjacency structure between adjacent cells. When using this type of matrix for subsequent windowing and feature extraction, the model can simultaneously perceive the spatial diffusion effect of local heat load changes and the thermal state of surrounding areas, avoiding the information fragmentation that occurs when relying solely on isolated site time-series predictions. This provides richer constraint information in the spatial dimension and effectively suppresses abnormal fluctuations in output caused by local data disturbances. The temporal statistical features and spatial neighborhood features of each training sample segment are extracted, and the training sample segments are classified according to these features to determine the load fluctuation pattern corresponding to each segment. The pattern is then used as a sample label and combined with the initial prediction model for iterative training. The temporal statistical features are composed of the peak, trough, rate of change, and dispersion of each dimension of the time series. The spatial neighborhood features are obtained by calculating the sum of the absolute values ​​of the differences between the central grid cell and all its adjacent grid cells at the same time. These two types of features are combined into a feature vector, which is then subjected to Euclidean distance measurement with the center vectors of multiple preset load fluctuation patterns to complete the classification from segment to pattern. Subsequent model training no longer directly uses the original data as the learning target, but instead uses the fluctuation pattern as supervisory information to guide parameter updates, prompting the prediction base model to learn the differentiated mapping patterns from input spatiotemporal data to heat load output under different fluctuation patterns. This approach enables the model to establish discrimination boundaries for modes such as steady, gradual increase, gradual decrease, sudden increase, and sudden decrease. When the input data exhibits early characteristics of a certain fluctuation pattern during real-time prediction, the model can output the predicted value according to the response logic of the corresponding mode. This effectively alleviates the output lag and overshoot problems of a single overall model when facing mode switching, and improves the prediction stability for extreme weather conditions and large-scale load changes. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0025] Figure 1 This is a flowchart of a smart heating load prediction method based on big data;

[0026] Figure 2 This is a flowchart of the standardized processing of multi-source heterogeneous time-series data;

[0027] Figure 3 This is a flowchart of the spatiotemporal correlation data matrix construction and training sample generation process;

[0028] Figure 4 This is a flowchart of the load forecasting basic model training and real-time data preprocessing process;

[0029] Figure 5 This is a flowchart of heat load prediction and online dynamic model correction. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] See Figure 1 This invention provides a smart heating load forecasting method and system based on big data. It includes: acquiring multi-source heterogeneous time-series data of a target area over a historical period, performing time alignment and outlier removal on the multi-source heterogeneous time-series data to generate standardized time-series data; mapping the standardized time-series data to corresponding spatial grid cells according to a preset spatial topology of the heating area to construct a spatiotemporal correlation data matrix; segmenting the spatiotemporal correlation data matrix using a sliding time window to generate multiple training sample segments; extracting the temporal statistical features and spatial neighborhood features of each training sample segment, and classifying the training sample segments according to the temporal statistical features and spatial neighborhood features to determine the load fluctuation pattern corresponding to each training sample segment; using the load fluctuation pattern as a sample label, iteratively training it in conjunction with a preset initial prediction model to obtain a basic load forecasting model; acquiring real-time weather forecast data and real-time heating system operation data of the target area, and generating real-time standardized data according to the steps of time alignment and outlier removal; and performing prediction calculations based on the real-time standardized data using the basic load forecasting model to output the predicted heat load value of the target area at a preset future time point.

[0032] Example 1:

[0033] In specific implementation, please refer to Figure 2 Multi-source heterogeneous time-series data should include at least outdoor temperature, outdoor humidity, wind speed, solar radiation intensity, return water temperature of each heating station, supply water pressure of each heating station, and flow rate of the heating network.

[0034] In some embodiments, the step of generating standardized time-series data is specifically implemented as follows: acquiring outdoor temperature, outdoor humidity, wind speed, and solar radiation intensity data of the target area collected by outdoor weather stations deployed within the target area during historical time periods, as well as return water temperature, supply water pressure, and heating network flow data of each heating station collected by sensors installed at monitoring nodes of each heating station and heating network. The above-mentioned data constitute multi-source heterogeneous time-series data, with differences in data acquisition frequency and timestamps from different data sources.

[0035] Based on a preset unified sampling frequency, linear interpolation is performed on data points with different timestamps in multi-source heterogeneous time-series data to align data points from all data sources to the same time axis. The preset unified sampling frequency is set according to the thermal inertia characteristics of the heating system and the performance indicators of the data acquisition equipment, ranging from 5 minutes / time to 15 minutes / time. In one embodiment, the preset unified sampling frequency is set to 10 minutes / time, based on the fact that a 10-minute time interval can capture the dynamic characteristics of heating load changes while avoiding data redundancy and wasted computing resources due to excessively high frequencies. During the linear interpolation process, for any two adjacent original timestamps, the estimated data value corresponding to the interpolation time point on the unified time axis is calculated. The linear interpolation formula is:

[0036]

[0037] in, Indicates the time point of interpolation Data valuation at the location, and These represent the timestamps of the two original data collections adjacent to the interpolation time point. Indicates the timestamp The raw data values ​​collected at the location, Indicates the timestamp The raw data values ​​collected at the location, This indicates the preset time point on a unified time axis where data valuation is required.

[0038] For each dimension of the aligned time-series data, the local moving average and local moving standard deviation are calculated. The local moving average is calculated using a sliding window with a width of a preset window length. The arithmetic mean of all data points within the window is calculated point-by-point along the time axis for each dimension of the time-series data. In one embodiment, the preset window length is set to 20 sampling points. If the uniform sampling frequency is 10 minutes / sample, the window covers a time span of 200 minutes. The local moving standard deviation is calculated based on the same sliding window, calculating the root mean square deviation of all data points within the window from the corresponding local moving average.

[0039] When the value of a data point exceeds the sum of the corresponding local moving average and the local moving standard deviation at a preset multiple threshold, the corresponding data point is marked as an outlier to be checked. The preset multiple threshold ranges from 2.5 to 4.0. In one embodiment, the preset multiple threshold is set to 3.0, based on the assumption that when the data distribution conforms to a normal distribution, the probability of normal data points falling within the range of plus or minus three standard deviations of the mean is higher than 99%, and data points exceeding this range can be judged as outliers. The marking process is executed independently for time series data of each dimension, ensuring that time series data of different physical dimensions are identified as outliers based on their own statistical distribution characteristics.

[0040] All marked outliers to be verified are removed from multi-source heterogeneous time-series data. Data is then filled into the removal location based on the average of the valid data points before and after the removal location. Specifically, for each removal location corresponding to an outlier, the nearest valid data point is searched forward and backward. The arithmetic mean of the two valid data points is calculated and filled into the removal location to generate standardized time-series data. If the removal location is at the end of the time-series data and valid data points in both directions cannot be obtained simultaneously, the value of the single valid data point closest to the removal location is copied and filled.

[0041] Example 2:

[0042] In specific implementation, please refer to Figure 3 Based on the preset spatial topology of the heating area, standardized time-series data are mapped to the corresponding spatial grid cells to construct a spatiotemporal correlation data matrix.

[0043] The steps for resolving the spatial topology of a heating area are as follows: First, acquire geographic information system (GIS) data of the heating network in the target area. Then, extract the three-dimensional geographic coordinates of all heating pipeline nodes, the geographic coordinates of all heating stations, and the topological connections between heating pipeline nodes and between heating pipeline nodes and heating stations from the GIS data. The three-dimensional geographic coordinates of the heating pipeline nodes include longitude, latitude, and altitude coordinates. The geographic coordinates of the heating stations include the longitude and latitude coordinates of the station's center point. The topological connections are stored in the form of an adjacency list. Each entry in the adjacency list records a node identifier and the set of all neighboring node identifiers directly connected to that node identifier. In one embodiment, the spatial topology of the heating area is mathematically represented using an undirected graph. Vertices in the undirected graph represent heating pipeline nodes and heating stations, and edges represent the physical pipeline connections between heating pipeline nodes or between heating pipeline nodes and heating stations.

[0044] Based on geographic coordinate information, the target area is divided into grids to determine the boundary range of each spatial grid unit. The gridding process adopts a uniform rectangular grid partitioning method, specifically: extracting the maximum and minimum longitude, maximum and minimum latitude values ​​from the geographic coordinate information of all heating pipeline nodes and all heating stations within the target area; determining the coverage span in the longitude direction based on the maximum and minimum longitude values, and determining the coverage span in the latitudinal direction based on the maximum and minimum latitude values. Spatial resolution parameters for the spatial grid units in the longitude and latitudinal directions are set. In one embodiment, the spatial resolution parameter in the longitude direction is set to 0.001 degrees, and the spatial resolution parameter in the latitudinal direction is set to 0.001 degrees. This setting is based on the fact that 0.001 degrees corresponds to a surface distance of approximately 111 meters, a distance scale that can distinguish the spatial differences in the areas managed by different heating stations. The coverage span in the longitude direction is divided into multiple equal parts according to the spatial resolution parameter in the longitude direction, and the coverage span in the latitudinal direction is divided into multiple equal parts according to the spatial resolution parameter in the latitudinal direction, forming a regular rectangular grid array. In a grid array, each rectangular cell serves as a spatial grid unit. The boundaries of each spatial grid unit are determined by four parameters: the longitude of the left boundary, the longitude of the right boundary, the latitude of the lower boundary, and the latitude of the upper boundary. Each spatial grid unit is assigned a unique grid index number, calculated using the following formula:

[0045]

[0046] in, The grid index number represents the spatial grid cell. This indicates the total number of columns of spatial grid cells divided along the longitude direction. This indicates the row number of the spatial grid cell, counted starting from the lower boundary in the latitudinal direction. This indicates the column number of the spatial grid cell, counted starting from the left boundary in the longitude direction. It is obtained by dividing the coverage span along the longitude direction by the spatial resolution parameter along the longitude direction and then rounding up. The value is a positive integer starting from 1. The value is a positive integer starting from 1.

[0047] Each data point in the standardized time-series data is assigned to a corresponding spatial grid cell based on its associated geographic location information. The geographic location information associated with each data point is determined by the longitude and latitude coordinates of the data acquisition device's deployment location. The process involves determining whether a data point's longitude coordinates fall between the left and right boundary longitude values ​​of a spatial grid cell, and whether its latitude coordinates simultaneously fall between the lower and upper boundary latitude values ​​of that spatial grid cell. If both conditions are met, the data point is assigned to that spatial grid cell. This determination and assignment process is repeated until all data points in the standardized time-series data are assigned to their respective spatial grid cells. In the specific implementation, a spatial hash mapping method is used to accelerate the assignment process. The longitude and latitude coordinates are discretized, and a hash key value is calculated. This hash key value is then used to directly locate the corresponding spatial grid cell.

[0048] For each spatial grid cell, the data values ​​of all data points contained within the spatial grid cell are arranged according to the time dimension to generate the corresponding time series vector for the spatial grid cell. When a spatial grid cell is assigned multiple data points at a timestamp on a unified time axis, the arithmetic mean of the data values ​​of these multiple data points is calculated, and this arithmetic mean is used as the data value for that spatial grid cell at that timestamp. When a spatial grid cell is not assigned any data points at a timestamp on a unified time axis, a linear interpolation method consistent with the steps for generating standardized time series data is used. The interpolation result is calculated using the valid data values ​​of the same spatial grid cell at adjacent timestamps, and the interpolation result is filled into the missing timestamp position. After the arrangement is completed, the time series vector of each spatial grid cell contains the data values ​​corresponding to all timestamps on the unified time axis, and the dimension of the time series vector is equal to the total number of timestamps on the unified time axis. The spatial grid cells contain data on outdoor temperature, outdoor humidity, wind speed, solar radiation intensity, return water temperature of each heating station, supply water pressure of each heating station, and flow rate of the heating network. Therefore, each spatial grid cell corresponds to a multidimensional feature vector at a single timestamp, and the number of dimensions of the multidimensional feature vector is equal to the total number of data types. The time series vector is essentially a two-dimensional matrix formed by arranging the multidimensional feature vectors along the time dimension.

[0049] The time-series vectors corresponding to all spatial grid cells are combined and arranged according to spatial topological adjacency relationships to form a spatiotemporal correlation data matrix. The spatial topological adjacency relationships are determined by the connection relationships recorded in the spatial topology of the heating area. The combination and arrangement method is as follows: for all spatial grid cells, a three-dimensional numerical array is constructed. The first dimension of the three-dimensional numerical array is the time dimension, and the second and third dimensions correspond to the row and column positions of the spatial grid cells in the grid array. The indexing method is as follows: the index values ​​of the second and third dimensions in the three-dimensional numerical array correspond to the row and column numbers of the corresponding spatial grid cells, and the index value of the first dimension corresponds to the timestamp sequence number on a unified time coordinate axis. The element corresponding to the position in the three-dimensional numerical array stores the multi-dimensional feature vector of the corresponding spatial grid cell at the corresponding timestamp. The resulting three-dimensional numerical array is the spatiotemporal correlation data matrix, which completely preserves the temporal evolution and spatial distribution information of heating-related data within the target area.

[0050] The specific implementation of the step of generating multiple training sample segments by segmenting the spatiotemporal correlation data matrix through a sliding time window is as follows: The length of the sliding time window and the sliding step size are set, with the length of the sliding time window being greater than the sliding step size. The length of the sliding time window is determined based on the minimum resolvable time span of the load fluctuation pattern and the required input time range of the prediction model, ranging from 24 to 168 time steps. In one embodiment, the length of the sliding time window is set to 72 time steps. When the uniform sampling frequency is 10 minutes / time, 72 time steps correspond to a time span of 12 hours. This is based on the premise that a 12-hour time window can completely capture the complete cycle of daytime and nighttime load alternation during the heating season. The sliding step size is determined based on the balance between the diversity of training samples and computational resources, ranging from 6 to 24 time steps. In one embodiment, the sliding step size is set to 12 time steps, corresponding to a time interval of 2 hours. This is based on the premise that there is a 10-hour time overlap between two adjacent training sample segments, avoiding the loss of load fluctuation patterns due to time offset.

[0051] Starting from the initial time point of the spatiotemporal correlation data matrix, the sliding time window moves progressively according to the sliding step size. The initial time point corresponds to the first timestamp position of the first time dimension in the spatiotemporal correlation data matrix. The movement process of the sliding time window is as follows: Define a window start position variable and a window end position variable. Initially, the window start position variable is assigned the index value of the initial time point, and the window end position variable is assigned the index value of the initial time point plus the sliding time window length minus one. After completing the data truncation at the current window position, the window start position variable is increased by the index offset corresponding to the sliding step size, and the window end position variable is simultaneously increased by the index offset corresponding to the sliding step size, thus obtaining the window position range corresponding to the next truncation operation.

[0052] After each move, a submatrix of all data within the current time window is extracted from the spatiotemporal correlation data matrix, and this extracted submatrix is ​​used as a training sample fragment. The extraction operation targets all values ​​of all spatial grid cells in the second and third dimensions of the spatiotemporal correlation data matrix within the time range covered by the current sliding time window, based on the multidimensional feature vectors. The size of the extracted submatrix in the time dimension is equal to the length of the sliding time window, its size in the two-dimensional spatial dimension is equal to the number of rows and columns of the grid array, and its size in the feature dimension is equal to the total number of data types.

[0053] Repeat the sliding window movement and data extraction process described above until the end of the sliding time window exceeds the time endpoint of the spatiotemporal correlation data matrix. The time endpoint corresponds to the last timestamp position of the first time dimension in the spatiotemporal correlation data matrix. The condition for determining that the end of the sliding time window has been exceeded is that the index value of the window end position variable is greater than the index value of the time endpoint. After the condition is met, stop the sliding window movement operation, and summarize all the data sub-matrices generated during the extraction process to obtain multiple training sample fragments.

[0054] Example 3:

[0055] In practice, the temporal statistical features and spatial neighborhood features of each training sample segment are extracted, and the training sample segments are classified according to the temporal statistical features and spatial neighborhood features to determine the load fluctuation pattern corresponding to each training sample segment.

[0056] For each training sample segment, calculate the peak, trough, rate of change, and dispersion of each dimension of the time series, and combine the calculation results to form time series statistical features.

[0057] Each dimension of a time series refers to a one-dimensional numerical sequence of individual spatial grid cells extracted from each training sample segment, arranged along the time dimension for a single data type. The total number of spatial grid cells in a single training sample segment is the product of the total number of columns of spatial grid cells along the longitude direction and the total number of rows of spatial grid cells along the latitude direction. The total number of data types in a single training sample segment is consistent with the total number of data types in the standardized time series data. Therefore, the total number of time series in each dimension is equal to the total number of spatial grid cells multiplied by the total number of data types.

[0058] Peak values ​​are obtained by iterating through all time steps in each dimension of the time series and recording the largest value encountered during the iteration. Valley values ​​are obtained by iterating through all time steps in each dimension of the time series and recording the smallest value encountered during the iteration.

[0059] The rate of change is calculated as follows: For each dimension of the time series, the absolute value of the difference between data values ​​between two adjacent time steps is calculated sequentially along the time axis. The absolute values ​​of the differences obtained from all adjacent time steps are added together and divided by the total number of time steps minus one to obtain the average absolute change. The average absolute change is then divided by the actual physical time interval between two adjacent time steps to obtain the rate of change. The actual physical time interval is determined by the uniform sampling frequency. When the uniform sampling frequency is set to 10 minutes / sample, the actual physical time interval is 10 minutes.

[0060] The degree of dispersion is measured by calculating the variance of the time series for each dimension. The variance is calculated as follows: calculate the arithmetic mean of all time-step data values ​​for each dimension; square the difference between each time-step data value and the arithmetic mean; sum all the squared values ​​and divide by the total number of time steps to obtain the variance. Alternatively, the degree of dispersion can be measured by calculating the standard deviation of the time series for each dimension, which is calculated by taking the square root of the variance.

[0061] The four statistical measures—peak, trough, rate of change, and dispersion—obtained from each dimension of the time series are concatenated in a fixed order to form the time series statistical feature sub-vector corresponding to each dimension of the time series. Then, the time series statistical feature sub-vectors corresponding to all time series of each dimension in a single training sample segment are concatenated sequentially, first in spatial grid cell index order, then in data type index order, to form the time series statistical features corresponding to that single training sample segment.

[0062] For each training sample segment, the sum of the absolute values ​​of the differences between the data values ​​of the central grid cell and all its neighboring grid cells at the same time is calculated, and the result is used as the spatial neighborhood feature.

[0063] The central grid cell is determined as follows: based on the spatial topology of the heating area, a spatial grid cell located at the load center of the heating network is selected from all spatial grid cells covered by a single training sample segment. Optionally, the central grid cell is selected as the spatial grid cell containing the largest number of heating stations. If multiple spatial grid cells with the same number of heating stations exist, the corresponding spatial grid cell closest to the geometric center of the target area is selected.

[0064] All adjacent grid cells refer to all spatial grid cells that are directly connected to the central grid cell in the spatial topological adjacency relationship. The determination of adjacent grid cells is based on the undirected graph constructed when analyzing the spatial topology of the heating area in Example 2. The spatial grid cells represented by all other vertices in the undirected graph that are directly connected to the vertices corresponding to the central grid cell are the adjacent grid cells.

[0065] The calculation method for data value differences at the same time step is as follows: For each time step in the time dimension of the training sample segment, obtain the multidimensional feature vectors corresponding to all data types of the central grid cell at that time step, and simultaneously obtain the multidimensional feature vectors corresponding to all data types of an adjacent grid cell at that time step. Calculate the Euclidean distance between the two multidimensional feature vectors; the resulting Euclidean distance value is the data value difference between the adjacent grid cell and the central grid cell at that time step. Sum the data value differences corresponding to all adjacent grid cells at a single time step to obtain the spatial neighborhood difference sum at that single time step. Sum the spatial neighborhood differences sums of all time steps in the time dimension of the training sample segment to obtain the spatial neighborhood features of that training sample segment.

[0066] The temporal statistical features and spatial neighborhood features are combined into a feature vector. Specifically, the spatial neighborhood features of a single training sample segment are added as an additional dimension and concatenated to the end of the temporal statistical features of that training sample segment to form a combined feature vector.

[0067] Calculate the Euclidean distance between the combined characteristic vector and the center vectors of multiple preset load fluctuation modes. These load fluctuation modes include a stable mode, a gradual increase mode, a gradual decrease mode, a sudden increase mode, and a sudden decrease mode.

[0068] The pre-defined method for obtaining the center vectors of multiple load fluctuation modes is as follows: Typical historical heating operation data covering all five load fluctuation modes over multiple past heating seasons in the target area are collected in advance. Following the same steps as generating training sample segments, the typical historical heating operation data is processed into multiple standard mode sample segments. For each standard mode sample segment, a combined feature vector is extracted using the same method described above. All combined feature vectors labeled with the same load fluctuation mode are subjected to cluster analysis. The K-means clustering algorithm is used to calculate the cluster centers, with K set to 1. The output cluster center vector is the center vector for that load fluctuation mode. Center vectors are generated for each of the five load fluctuation modes, and the dimension of the center vector is the same as the dimension of the combined feature vector.

[0069] The load fluctuation characteristics of the stable mode are: the heat load value fluctuates slightly around a stable average value within the time window, without a significant upward or downward trend. The load fluctuation characteristics of the gradual increase mode are: the heat load value shows a slow increasing trend within the time window, with the increase between adjacent time steps consistently positive and the absolute value of the increase being small. The load fluctuation characteristics of the gradual decrease mode are: the heat load value shows a slow decreasing trend within the time window, with the decrease between adjacent time steps consistently positive and the absolute value of the decrease being small. The load fluctuation characteristics of the sudden increase mode are: the heat load value experiences a significant jump within a short time interval within the time window, with the jump significantly greater than the changes in other time periods. The load fluctuation characteristics of the sudden decrease mode are: the heat load value experiences a significant drop within a short time interval within the time window, with the drop significantly greater than the changes in other time periods.

[0070] The formula for calculating Euclidean distance is:

[0071]

[0072] in, Represents the combination of eigenvectors and the first The Euclidean distance between the center vectors of the load fluctuation patterns The values ​​cover the corresponding numbers for five load fluctuation modes: steady mode, gradual increase mode, gradual decrease mode, sudden increase mode, and sudden decrease mode. This represents the total dimension of the combined feature vectors. Represents the first eigenvector in the combined feature vector. The element values ​​corresponding to each dimension Indicates the first The first load fluctuation mode center vector The element values ​​corresponding to each dimension are obtained by calculating the square root of the sum of the squared differences between the combined feature vector and the load fluctuation pattern center vector in all dimensions.

[0073] The load fluctuation pattern with the smallest Euclidean distance is determined as the load fluctuation pattern corresponding to the training sample segment. In the specific implementation, the Euclidean distance is calculated for each of the five load fluctuation patterns. Five distance values ​​are obtained. The magnitudes of the five distance values ​​are compared, and the load fluctuation pattern corresponding to the smallest distance value is selected as the classification result of the training sample segment. The classification result is recorded as the load fluctuation pattern corresponding to the training sample segment.

[0074] Example 4:

[0075] In specific implementation, please refer to Figure 4 By using load fluctuation patterns as sample labels and combining them with a pre-set initial prediction model for iterative training, a basic load prediction model is obtained.

[0076] The core architecture of the preset initial prediction model adopts a spatiotemporal convolutional recurrent neural network, consisting of a spatial feature extraction module, a temporal feature extraction module, and a pattern classification module connected sequentially. The spatial feature extraction module contains two 2D convolutional layers. The first 2D convolutional layer has 32 kernels of size 3×3, a stride of 1, equal-size edge padding, and uses a linear rectified function as the activation function. The second 2D convolutional layer has 64 kernels of size 3×3, a stride of 1, equal-size edge padding, and uses a linear rectified function as the activation function. A batch normalization layer and a 2D max-pooling layer are inserted between the two 2D convolutional layers. The 2D max-pooling layer has a pooling window size of 2×2 and a pooling stride of 2. The temporal feature extraction module contains a two-layer long short-term memory network. The first long short-term memory network has 128 hidden units and outputs the hidden states of all time steps in the sequence. The second long short-term memory network has 64 hidden units and outputs only the hidden state of the last time step. The pattern classification module includes a fully connected layer and a flexible maximum output layer. The number of neurons in the fully connected layer is set to 32, and the activation function is a linear rectified function. The number of neurons in the flexible maximum output layer is set to 5, corresponding one-to-one with the five load fluctuation patterns, and outputting the predicted probability distribution of the five load fluctuation patterns.

[0077] The training sample segments are used as input data, and the corresponding load fluctuation patterns are used as supervision information, both input into a pre-defined initial prediction model. Before being input into the pre-defined initial prediction model, the training sample segments undergo dimensionality rearrangement. After dimensionality rearrangement, the tensor shape of the training sample segment is [number of time steps, number of spatial grid cell rows, number of spatial grid cell columns, number of data type channels]. The number of time steps equals the length of the sliding time window, the number of spatial grid cell rows equals the total number of spatial grid cell rows in the latitudinal direction, the number of spatial grid cell columns equals the total number of spatial grid cell columns in the longitudinal direction, and the number of data type channels equals the total number of all data types involved in the standardized time series data. The supervision information is represented using one-hot encoding. The one-hot encoded vector has a length of 5, with a 1 at one position indicating that the training sample segment belongs to the corresponding load fluctuation pattern, and 0 at other positions.

[0078] Based on the current parameters of the preset initial prediction model, the cross-entropy loss value between the predicted fluctuation pattern output by the preset initial prediction model and the supervision information is calculated. The formula for calculating the cross-entropy loss value is:

[0079]

[0080] in, This represents the cross-entropy loss value. This indicates the total number of categories of load fluctuation patterns. The value of is fixed at 5. This represents the first one-hot encoded vector of supervised information. The value of each element, This represents the output of the first layer of the preset initial prediction model, which is the maximum flexibility value. The predicted probability value of the load fluctuation pattern. The value of is 0 or 1, if and only if the training sample segment is labeled as the . When the load fluctuation mode The value is 1. For the preset initial prediction model, the first The prediction confidence level for the load fluctuation pattern is a real number ranging from 0 to 1.

[0081] Based on the cross-entropy loss value, the parameters of the preset initial prediction model are updated via backpropagation using gradient descent. The gradient descent method employs an adaptive momentum estimation optimizer, with the learning rate parameter set to 0.001, the first-order momentum decay coefficient β1 set to 0.9, the second-order momentum decay coefficient β2 set to 0.999, and the numerical stability constant ε set to... During the backpropagation update process, the partial derivatives of the cross-entropy loss value with respect to all trainable parameters of the preset initial prediction model are calculated. The adaptive momentum estimation optimizer is then used to calculate the adaptive learning rate adjustment for each trainable parameter based on the partial derivatives. The updated parameter value is obtained by subtracting the product of the adaptive learning rate adjustment and the learning rate parameter from the current parameter value.

[0082] Repeat the above input, calculation, and update process, with each complete input, calculation, and update process recorded as one iteration. After each iteration, record the cross-entropy loss value. When the change in the cross-entropy loss value over 50 consecutive iterations is less than a preset convergence threshold, training stops, and the model with the current parameters is used as the base model for load prediction. The preset convergence threshold is set to 0.0001, based on the premise that when the decrease in the loss value is less than 0.0001, the model parameters are close to a local optimum, and the performance improvement from continued training is negligible, while also avoiding overfitting.

[0083] In some embodiments, real-time weather forecast data and real-time heating system operation data for the target area are acquired, and real-time standardized data are generated by following the steps of time alignment and outlier removal.

[0084] The system receives meteorological forecast data for the target area at multiple points in advance, as well as real-time operational data from various monitoring points of the heating system. The meteorological forecast data originates from a numerical weather prediction system and includes gridded forecast products covering the target area. The data includes at least forecast values ​​for outdoor temperature, outdoor humidity, wind speed, and solar radiation intensity. The real-time operational data comes from sensors deployed at various monitoring points of the heating system and includes at least real-time values ​​for return water temperature, supply water pressure, and flow rate in the heating network at each heating station. The temporal resolution of the meteorological forecast data is consistent with the output frequency of the numerical weather prediction system, and the acquisition frequency of the real-time operational data is consistent with the polling cycle of the heating system's data acquisition and monitoring control system.

[0085] Time alignment is performed on weather forecast data and real-time operational data to synchronize them to the same time baseline. The time alignment operation is consistent with the time alignment method used in the step of generating standardized time series data. Based on a unified sampling frequency, linear interpolation is performed on data points with different timestamps in the weather forecast data and real-time operational data. The linear interpolation formula is the same as the formula used in the step of generating standardized time series data, and the unified sampling frequency remains unchanged at 10 minutes / time.

[0086] For the aligned real-time data, the same methods for calculating the local moving average and local moving standard deviation are used to identify and remove outliers. The sliding window width for the local moving average is set to 20 sampling points, and the local moving standard deviation is calculated based on the same sliding window, with a preset multiplier threshold of 3.0. Each time-series data point in each dimension of the real-time data is compared point-by-point with the sum of the local moving average and three times the local moving standard deviation. When the value of a data point exceeds this sum, the data point is marked as a real-time outlier and removed from the real-time data.

[0087] The removal location is filled using the average of the valid real-time data points before and after it. For each real-time outlier corresponding to a removal location, the nearest valid real-time data point is searched forward and backward. The arithmetic mean of the values ​​of the two valid real-time data points is calculated and filled into the removal location. If the removal location is located at the end of the real-time data time series, the value of the nearest valid real-time data point is used for filling. After filling, real-time standardized data is generated.

[0088] Example 5:

[0089] In specific implementation, please refer to Figure 5 Based on real-time standardized data, the system performs prediction calculations using a load forecasting model and outputs the predicted heat load value for the target area at a future preset time point.

[0090] In some embodiments, real-time standardized data is mapped to spatial grid cells according to the spatial topology of the heating area to generate a real-time spatiotemporal correlation data matrix. The real-time standardized data includes all data received at multiple times prior to the current time point, after time alignment and outlier removal processing of meteorological forecast data and real-time operation data of the heating system. The geographical location information associated with each data point remains consistent with that of the training phase. The mapping process is implemented in the same way as the method of constructing the spatiotemporal correlation data matrix in Example 2: Parse the geographic information system data of the heating pipeline network in the target area to obtain the geographic coordinate information of the heating pipeline nodes and heating stations; divide the target area into uniform rectangular grids based on a spatial resolution parameter of 0.001 degrees in the longitude direction and 0.001 degrees in the latitude direction, and determine the boundary range of each spatial grid unit; compare the longitude and latitude coordinates of each data point in the real-time standardized data with the boundary of the spatial grid unit, and assign each data point in the real-time standardized data to its respective spatial grid unit; calculate the arithmetic mean of the data values ​​of all data points in each spatial grid unit at each unified sampling timestamp to form a multidimensional feature vector of the corresponding spatial grid unit, and arrange the multidimensional feature vectors along the time dimension to generate a time series vector; combine the time series vectors of all spatial grid units into a three-dimensional numerical array according to the spatial topological adjacency relationship to obtain the real-time spatiotemporal correlation data matrix.

[0091] A sliding time window is used to extract predicted input segments from the real-time spatiotemporal correlation data matrix. The length of the sliding time window is set to the same value as the training sample segment generation process, i.e., 72 time steps. The sliding step size no longer affects the extraction process, and the extraction position is fixed at the data submatrix of the 72 consecutive time steps closest to the current time point in the real-time spatiotemporal correlation data matrix. The extraction operation traces backward from the end of the time dimension of the real-time spatiotemporal correlation data matrix back to all time steps covered by the length of the sliding time window, extracting the values ​​corresponding to all data types of all spatial grid cells at these time steps to form the predicted input segment. The tensor shape of the predicted input segment is [72, number of rows of spatial grid cells, number of columns of spatial grid cells, number of data type channels]. The number of data type channels is equal to the total number of data types of outdoor temperature, outdoor humidity, wind speed, solar radiation intensity, return water temperature of each heating station, supply water pressure of each heating station, and flow rate of the heating network.

[0092] The predicted input fragments are fed into the load forecasting base model, driving its forward propagation calculation process. The load forecasting base model's architecture is based on the preset initial prediction model described in Example 4, with the addition of a heat load regression module. The load forecasting base model consists of a spatial feature extraction module, a temporal feature extraction module, a pattern classification module, and a heat load regression module. The spatial feature extraction module contains two two-dimensional convolutional layers. The first two-dimensional convolutional layer uses 32 3×3 convolutional kernels with a stride of 1, equal-size edge padding, and a linear rectified function as the activation function. The second two-dimensional convolutional layer uses 64 3×3 convolutional kernels with a stride of 1, equal-size edge padding, and a linear rectified function as the activation function. A batch normalization layer and a two-dimensional max-pooling layer with a pooling window size of 2×2 and a pooling stride of 2 are connected in series between the two layers. The temporal feature extraction module comprises a two-layer Long Short-Term Memory (LSTM) network. The first LSM layer has 128 hidden units and outputs the hidden states for all time steps. The second LSM layer has 64 hidden units and outputs only the hidden state for the last time step. The pattern classification module receives the hidden state of the last time step from the second LSM layer, passes it through a fully connected layer with 32 neurons and a flexible maximum output layer with 5 neurons, and outputs the predicted probabilities of five load fluctuation patterns. The heat load regression module is connected in parallel with the pattern classification module after the temporal feature extraction module. It also receives the hidden state of the last time step from the second LSM layer, passes it through a fully connected layer with 64 neurons and a linear activation output layer, and outputs a dimension equal to the total number of spatial grid cells. Each output value represents the predicted heat load value of the corresponding spatial grid cell at a preset future time point. While training using load fluctuation patterns as sample labels as described in Example 4, the actual heat load values ​​of each spatial grid cell at future times corresponding to the training sample segments are used as the supervision signal of the heat load regression module. A weighted sum of cross-entropy loss and mean squared error loss is jointly calculated, with the cross-entropy loss weight coefficient set to 1.0 and the mean squared error loss weight coefficient set to 0.5. Joint backpropagation updates are performed using an adaptive momentum estimation optimizer with a learning rate parameter of 0.001, a first-order momentum decay coefficient of 0.9, a second-order momentum decay coefficient of 0.999, and a numerical stability constant of 10⁻⁻⁶. 8 After training and convergence, a basic load forecasting model with complete predictive capabilities is obtained. During forecasting calculation, the input segments are sequentially passed through the spatial feature extraction module and the temporal feature extraction module. The hidden state of the last time step output by the temporal feature extraction module is simultaneously passed to the pattern classification module and the heat load regression module, directly obtaining the result of the linear activation output layer of the heat load regression module.

[0093] Obtain the results from the output layer of the load forecasting base model and determine these results as the predicted heat load values ​​for each spatial grid cell corresponding to a future preset time point in the target area. The a-th element in the numerical vector output by the linear activation output layer corresponds to the predicted heat load value of the spatial grid cell with grid index number a.

[0094] In some embodiments, a model dynamic correction step is performed after the predicted heat load value is generated.

[0095] Record the actual heat load value of the target area at a future preset time point. The actual heat load value is obtained as follows: after the future preset time point arrives, read the measured heat load value of the heat meters deployed in the corresponding area of ​​each spatial grid unit at that time point from the heating system data acquisition and monitoring control system. For each spatial grid unit, sum the measured heat load values ​​of all heat meters within the spatial grid unit to obtain the actual heat load value of the spatial grid unit. The actual heat load values ​​of all spatial grid units constitute the set of actual heat load values.

[0096] Calculate the prediction deviation between the actual heat load and the predicted heat load. Sum the predicted heat load values ​​of all spatial grid cells within the target area to obtain the total predicted heat load value for the target area; sum the actual heat load values ​​of all spatial grid cells to obtain the total actual heat load value for the target area. The formula for calculating the prediction deviation is:

[0097]

[0098] in, This represents the prediction deviation value. This represents the actual total heat load value of the target area, obtained by summing the actual heat load values ​​of all spatial grid cells at a future preset time point. This represents the predicted total heat load value for the target area, obtained by summing the predicted heat load values ​​of all spatial grid cells output by the load prediction basic model. This indicates the operation of taking the absolute value.

[0099] When the absolute value of the prediction deviation exceeds a preset deviation tolerance threshold multiple times consecutively, an online learning mechanism is triggered. The preset deviation tolerance threshold is determined based on the design capacity of the target area's heating system and the historical heat load fluctuation range, and is set to 5% of the target area's design heat load capacity. In one embodiment, the number of consecutive occurrences is set to 3. Each prediction deviation value is recorded. When it is detected that the three most recent consecutive prediction deviation values ​​are all greater than the preset deviation tolerance threshold, it is determined that the prediction deviation value has exceeded the preset deviation tolerance threshold multiple times consecutively, triggering the online learning mechanism.

[0100] Under the online learning mechanism, the latest data containing actual heat load values ​​is added to the historical dataset, and the time alignment, outlier removal, and model training steps are re-executed to update the parameters of the load prediction base model. The latest data collection scope includes all real-time weather forecast data, real-time heating system operation data, and actual heat load values ​​at corresponding future preset time points within a time span of one sliding time window from the time the online learning mechanism is triggered. The latest data is merged with the original historical dataset. The merged dataset is then subjected to the time alignment operation described in Example 1, which involves linear interpolation at a uniform sampling frequency of 10 minutes / time. The local moving average and local moving standard deviation are calculated, and outliers are marked and removed using a preset multiple threshold of 3.0. The updated standardized time series data is generated by filling in the mean of the effective data points before and after the merged dataset. The steps described in Example 2, namely, constructing a spatiotemporal correlation data matrix and extracting training sample segments using a sliding time window, are then performed again. The steps described in Example 3, namely, extracting time series statistical features and spatial neighborhood features, calculating the Euclidean distance between the combined feature vector and the center vectors of the five load fluctuation patterns, and determining the load fluctuation pattern, are then performed again. The training steps described in Example 4, namely, inputting the training sample segments and load fluctuation pattern labels into the model, calculating the weighted sum of cross-entropy loss and mean squared error loss, and backpropagating to update the parameters, are then performed again. The training convergence threshold remains unchanged, resulting in updated load forecasting basic model parameters. The updated parameters are then used to overwrite the parameters of the original load forecasting basic model.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart heating load forecasting method based on big data, characterized in that, The method includes: Acquire multi-source heterogeneous time-series data of the target region within a historical time period, and perform time alignment and outlier removal on the multi-source heterogeneous time-series data to generate standardized time-series data; Based on the preset spatial topology of the heating area, the standardized time-series data is mapped to the corresponding spatial grid cells to construct a spatiotemporal correlation data matrix; Multiple training sample segments are generated by segmenting the spatiotemporal correlation data matrix using a sliding time window. Extract the temporal statistical features and spatial neighborhood features of each training sample segment, and classify the training sample segments according to the temporal statistical features and spatial neighborhood features to determine the load fluctuation pattern corresponding to each training sample segment; The load fluctuation pattern is used as a sample label, and iterative training is performed in combination with a preset initial prediction model to obtain a basic load prediction model. Acquire real-time weather forecast data and real-time heating system operation data for the target area, and generate real-time standardized data according to the time alignment and outlier removal steps. Based on the real-time standardized data, the load prediction model is used to perform prediction calculations and output the predicted heat load value of the target area at a future preset time point.

2. The intelligent heating load prediction method based on big data according to claim 1, characterized in that, The multi-source heterogeneous time-series data includes at least outdoor temperature, outdoor humidity, wind speed, solar radiation intensity, return water temperature of each heating station, supply water pressure of each heating station, and flow rate of the heating network.

3. The intelligent heating load prediction method based on big data according to claim 1, characterized in that, The steps of acquiring multi-source heterogeneous time-series data of the target region within a historical time period, and performing time alignment and outlier removal on the multi-source heterogeneous time-series data to generate standardized time-series data specifically include: Based on a preset uniform sampling frequency, linear interpolation is performed on data points with different timestamps in the multi-source heterogeneous time series data so that data points from all data sources are aligned to the same time coordinate axis. For each dimension of the time series data after alignment, calculate its local moving average and local moving standard deviation; When the value of a data point exceeds the sum of its corresponding local moving average and the local moving standard deviation under a preset multiple threshold, the data point is marked as an outlier to be checked. The outliers to be checked are removed from the multi-source heterogeneous time series data, and the data at the removal position is filled with data based on the average of the valid data points before and after the removal position, thereby generating the standardized time series data.

4. The intelligent heating load prediction method based on big data according to claim 3, characterized in that, The step of mapping the standardized time-series data to corresponding spatial grid cells based on a preset spatial topology of the heating area to construct a spatiotemporal correlation data matrix specifically includes: The spatial topology of the heating area is analyzed to obtain the geographic coordinates of all heating pipeline nodes and heating stations within the target area; Based on the geographic coordinate information, the target area is divided into grids to determine the boundary range of each spatial grid unit; Each data point in the standardized time series data is assigned to a corresponding spatial grid cell according to its associated geographical location information; For each spatial grid cell, the data values ​​of all the data points contained within it are arranged according to the time dimension to generate the time series vector of that grid cell; The time series vectors corresponding to all spatial grid units are combined and arranged according to spatial topological adjacency relationships to form the spatiotemporal correlation data matrix.

5. The intelligent heating load prediction method based on big data according to claim 4, characterized in that, The step of segmenting the spatiotemporal correlation data matrix using a sliding time window to generate multiple training sample segments specifically includes: Set the length of the sliding time window and the sliding step size, wherein the length is greater than the sliding step size; Starting from the time start point of the spatiotemporal correlation data matrix, the sliding time window is moved step by step according to the sliding step size; After each move, all data submatrices located within the current time window are extracted from the spatiotemporal correlation data matrix, and this data submatric is used as a training sample fragment. Repeat the above extraction process until the end of the sliding time window exceeds the time endpoint of the spatiotemporal correlation data matrix, thereby generating the multiple training sample segments.

6. The intelligent heating load prediction method based on big data according to claim 5, characterized in that, The steps of extracting the temporal statistical features and spatial neighborhood features of each training sample segment, and classifying the training sample segments based on the temporal statistical features and spatial neighborhood features to determine the load fluctuation pattern corresponding to each training sample segment, specifically include: For each training sample segment, calculate the peak, trough, rate of change, and dispersion of each dimension of the time series, and combine the calculation results to form the time series statistical features. For each training sample segment, calculate the sum of the absolute values ​​of the differences between the data values ​​of its central grid cell and all its neighboring grid cells at the same time, and use the calculation result as the spatial neighborhood feature; The time-series statistical features and the spatial neighborhood features are combined into a feature vector, and the Euclidean distance between the feature vector and the preset center vectors of multiple load fluctuation patterns is calculated. The load fluctuation pattern with the smallest Euclidean distance is determined as the load fluctuation pattern corresponding to the training sample segment.

7. The intelligent heating load prediction method based on big data according to claim 6, characterized in that, The degree of dispersion is measured by calculating the variance or standard deviation of the time series for each dimension.

8. The intelligent heating load prediction method based on big data according to claim 6, characterized in that, The various load fluctuation modes include a stable mode, a gradual increase mode, a gradual decrease mode, a sudden increase mode, and a sudden decrease mode.

9. The intelligent heating load prediction method based on big data according to claim 6, characterized in that, The steps of using the load fluctuation pattern as sample labels and iteratively training it in conjunction with a preset initial prediction model to obtain the basic load prediction model specifically include: The training sample fragments are used as input data, and the corresponding load fluctuation patterns are used as supervision information, which are then input into the preset initial prediction model. Based on the current parameters of the preset initial prediction model, calculate the cross-entropy loss value between the predicted fluctuation pattern output by the model and the supervision information. Based on the cross-entropy loss value, the parameters of the preset initial prediction model are updated by backpropagation using the gradient descent method. Repeat the above input, calculation and update process until the change in the cross-entropy loss value in consecutive iterations is less than the preset convergence threshold, then stop training and use the model under the current parameters as the basic model for load prediction.

10. A smart heating load forecasting system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the smart heating load prediction method based on big data as described in any one of claims 1 to 9.

Images