A photoelectric-building group load space-time matching degree prediction system

Abstract

The application provides a photovoltaic-building load space-time matching degree prediction system, comprising: a perception layer for collecting photovoltaic output data and building group load data; a network layer for transmitting the collected data to a control layer; the control layer is used for constructing a photovoltaic-building load coupling model integrating time and space dimensions based on the photovoltaic output data and the building group load data, performing multi-time scale collaborative prediction, and calculating time matching degree, space matching degree and comprehensive space-time matching degree based on the prediction results; the application layer is used for visualizing and displaying the prediction results, the time matching degree, the space matching degree and the comprehensive space-time matching degree. By constructing the photovoltaic-building load coupling model integrating time and space dimensions and performing multi-scale collaborative prediction, the future state of photovoltaic output and building load can be accurately predicted, and the matching degree of the two in time and space can be quantitatively output, solving the defects of independent prediction and inability to evaluate the space-time matching state.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a photovoltaic-building cluster load spatiotemporal matching degree prediction system. Background Technology

[0002] The coupled application of photovoltaic power generation with integrated building energy systems is becoming increasingly widespread. Existing photovoltaic-building complex systems typically use independent photovoltaic output forecasting and building load forecasting methods when conducting supply and demand analysis. This results in an inability to accurately predict the future spatiotemporal matching of supply and demand, making it difficult to support forward-looking system scheduling decisions. Summary of the Invention

[0003] This invention provides a photovoltaic-building cluster load spatiotemporal matching degree prediction system to solve the defect of existing technologies that cannot accurately predict the future spatiotemporal matching state of supply and demand.

[0004] This invention provides a photovoltaic-building cluster load spatiotemporal matching degree prediction system, comprising: The sensing layer is used to collect photovoltaic power output data and building cluster load data. The network layer is used to transmit the photovoltaic power output data and the building cluster load data to the control layer; The control layer is used to construct a photovoltaic-building load coupling model that integrates time and space dimensions based on the photovoltaic power output data and the building group load data, perform multi-time scale collaborative prediction, and calculate the time matching degree, spatial matching degree and comprehensive spatiotemporal matching degree based on the prediction results. The application layer is used to visualize the prediction results, the temporal matching degree, the spatial matching degree, and the comprehensive spatiotemporal matching degree.

[0005] According to the present invention, a photovoltaic-building cluster load spatiotemporal matching degree prediction system is provided, wherein the control layer adopts a cloud-edge-device collaborative computing architecture, wherein: A cloud computing platform for running long-term historical data analysis and first-scale prediction; Edge computing nodes are deployed on the side of the building complex to perform real-time data cleaning and second-scale prediction; An end-side intelligent control device is used to perform third-scale prediction; Wherein, the duration of the first scale is greater than the duration of the second scale, and the duration of the second scale is greater than the duration of the third scale.

[0006] According to the photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention, the control layer, when executing the construction of a photoelectric-building load coupling model that integrates temporal and spatial dimensions, is specifically used for: Cluster analysis is performed on historical and real-time data to generate composite cluster labels that include weather conditions, personnel patterns, and equipment status. Using the composite cluster label as a condition, the cross-correlation function of the photovoltaic power output sequence and the building load sequence under different time delays is calculated to obtain the scenario-based time lag characteristics; By combining distribution network topology, geographic information, and cloud motion spatial propagation characteristics, an equivalent fusion distance representing the relationship between electrical connections and spatial propagation is constructed; Using a graph attention network, the composite cluster label, the time lag feature, and the equivalent fusion distance are taken as inputs to dynamically learn and output the spatiotemporal correlation weight matrix between photovoltaic nodes and load nodes, which serves as a photovoltaic-building load coupling model.

[0007] According to the photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention, the control layer is specifically used for: On a cloud computing platform, long short-term memory networks are used to perform first-scale predictions with the first prediction span and the first time granularity. In the edge computing node device, a Transformer encoder-decoder network is used to perform a second-scale prediction with a second prediction span and a second time granularity, and residual correction is performed with the result of the first-scale prediction as a baseline. On the edge intelligent control device, a lightweight model combining convolutional networks and gated recurrent units is used to perform third-scale predictions of the third prediction span and third time granularity. Wherein, the first prediction span is greater than the second prediction span, the second prediction span is greater than the third prediction span, the first time granularity is greater than the second time granularity, and the second time granularity is greater than the third time granularity.

[0008] According to the photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention, the control layer, after performing multi-timescale collaborative prediction, is further used for: Real-time observation data is introduced, and the results of the first-scale prediction, the second-scale prediction, and the third-scale prediction are rolled over and corrected using a Bayesian update framework. The rolling correction results are optimized by a weight fusion mechanism that combines time decay and error adaptation to generate the final predicted sequence of photovoltaic output and building load.

[0009] According to the photoelectric-building complex load spatiotemporal matching degree prediction system provided by the present invention, the control layer is specifically used for: calculating the temporal matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree. Within the calculation window, after time alignment of the photovoltaic output and building load power sequences in the final predicted sequence, the normalized cross-correlation coefficient is calculated as the time matching degree. Based on the equivalent fusion distance, the photovoltaic supply and load demand in the final predicted sequence are weighted by spatial attenuation weights to calculate the matching ratio, which is used as the spatial matching degree. The temporal matching degree and the spatial matching degree are weighted and summed to obtain a comprehensive spatiotemporal matching degree index.

[0010] According to the photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention, the control layer is further used for: By calling upon the physical models of building thermal storage bodies, energy storage devices, and adjustable loads, and combining them with real-time equipment status and indoor comfort constraints, a demand-side flexible adjustment potential curve is generated. Based on the comprehensive spatiotemporal matching degree and the flexible adjustment potential curve, a corresponding scheduling strategy is generated; Accordingly, the application layer is also used to visualize the flexible adjustment potential curve and the scheduling strategy.

[0011] According to the photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention, the control layer is specifically used for generating the demand-side flexible adjustment potential curve when: For each type of flexible load resource, the theoretical adjustable power, comfort sensitivity coefficient, and real-time availability coefficient are multiplied to obtain the effective adjustable power of the corresponding flexible load resource. The effective adjustable power of the flexible load resources is aggregated and optimized by combining the time-period electricity price signal, the comprehensive spatiotemporal matching degree, and equipment operation constraints to generate a demand-side flexible adjustment potential curve.

[0012] According to the present invention, a photoelectric-building cluster load spatiotemporal matching degree prediction system is provided, wherein the sensing layer includes: Irradiance meters, sky imaging cameras, module temperature sensors, and electrical measurement devices deployed on the photovoltaic side are used to collect irradiance, cloud motion characteristics, module temperature, and output power. Smart meters, heating and cooling meters, environmental sensors, and occupancy monitoring devices deployed on the side of the building complex are used to collect data on sub-items of electrical load, heating and cooling load, indoor environmental parameters, and occupancy status.

[0013] According to the present invention, a photoelectric-building cluster load spatiotemporal matching degree prediction system is provided, wherein the application layer is specifically used for: Real-time display of predicted and measured curves for photovoltaic power output and building load; The dynamic changes of the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree are visualized.

[0014] This invention provides a photovoltaic-building cluster load spatiotemporal matching degree prediction system, comprising: a perception layer for collecting photovoltaic output data and building cluster load data; a network layer for transmitting photovoltaic output data and building cluster load data to a control layer; a control layer for constructing a photovoltaic-building load coupling model integrating time and space dimensions based on the photovoltaic output data and building cluster load data, performing multi-timescale collaborative prediction, and calculating time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree based on the prediction results; and an application layer for visually displaying the prediction results and the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree. By constructing a photovoltaic-building load coupling model integrating spatiotemporal dimensions and performing multi-scale collaborative prediction, it is possible to accurately predict the future state of photovoltaic output and building load, and quantify the degree of matching between the two in time and space. This solves the defects of independent prediction and inability to assess spatiotemporal matching state, providing a core decision-making basis for the collaborative scheduling and optimized operation of the building cluster integrated energy system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of a photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention; Figure 2 This is a schematic diagram of the load subsequence generation process provided by the present invention; Figure 3 This is a schematic diagram of the data dimensionality reduction process provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Figure 1 This is a schematic diagram of the structure of a photoelectric-building cluster load spatiotemporal matching degree prediction system provided by the present invention.

[0019] like Figure 1As shown, to address the problem that existing solutions predict supply and demand independently without considering their spatiotemporal correlation and coupling effects, this embodiment provides a photovoltaic-building cluster load spatiotemporal matching degree prediction system, comprising: a perception layer 1, used to collect photovoltaic output data and building cluster load data; a network layer 2, used to transmit photovoltaic output data and building cluster load data to a control layer 3; a control layer 3, used to construct a photovoltaic-building load coupling model integrating time and space dimensions based on photovoltaic output data and building cluster load data, perform multi-timescale collaborative prediction, and calculate time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree based on the prediction results; and an application layer 4, used to visualize the prediction results and time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree.

[0020] Specifically, Sensing Layer 1 consists of a series of distributed physical sensing devices responsible for comprehensive status sensing and data acquisition of the photovoltaic power generation system and the building complex's energy consumption system. Ground-based irradiance meters are deployed to collect global horizontal irradiance (GHI), direct normal irradiance (DNI), and horizontal diffuse irradiance (DHI). Sky imaging cameras capture cloud distribution and cloud movement characteristics. Smart sensors monitor module temperature, string voltage / current, and inverter operating power in real time, achieving photovoltaic-side sensing. Smart meters and heat meters are used to achieve separate metering of HVAC, lighting sockets, and power equipment. Environmental sensors are deployed to monitor indoor temperature and humidity. Concentration, combined with access control / camera / infrared sensor statistics on personnel activity and occupancy status, enables building-wide perception.

[0021] Network layer 2 constructs a highly reliable, low-latency data transmission channel, connecting physical sensing devices with the computing / control center. This enables data communication and connection between sensing layer 1 and control layer 3, including a wireless sensor network, an IoT gateway, and meteorological data access. The wireless sensor network utilizes self-organizing networks, LoRaWAN long-distance transmission, and 5G technology. The IoT gateway includes Modbus RTU / TCP, BACnet / TP, CANopen, and RS485. Meteorological data access includes temperature, humidity, air pressure, wind speed, and wind direction. Communication supports industrial private networks, 5G communication, and wired distribution network communication technologies, ensuring real-time uploading of multi-source heterogeneous data. Data interaction enables millisecond-level edge-side data acquisition and minute-level cloud data aggregation, while also accessing external meteorological forecast data to provide environmental input for prediction models.

[0022] Control Layer 3 is responsible for the entire closed-loop process from data processing and modeling prediction to strategy generation. It primarily processes photovoltaic (PV) power output data and building load data, establishing a PV-building load coupling model to perform multi-timescale collaborative prediction. These timescales include a first scale (24 hours), a second scale (4 hours), and a third scale (15 minutes). After completing the multi-scale collaborative prediction, it performs spatiotemporal correlation analysis based on the results to determine the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree. By performing interpretable matching degree calculations on the "supply side (PV power output) - demand side (building load)" under a unified time benchmark, it outputs a set of matching degree indicators that can be directly used for strategy judgment and dispatchable potential calculation. The time matching degree indicator is used to assess temporal consistency, the spatial matching degree indicator is used to assess spatial coupling consistency, and the comprehensive matching degree indicator is used to assess the degree of comprehensive spatiotemporal matching degree. This avoids misjudgments caused by static measurements based solely on power differences and ensures the engineering usability of the indicators under multiple scenarios and spatial constraints.

[0023] The control layer 3 adopts a cloud-edge-device collaborative computing architecture, where: a cloud computing platform is used to run long-term historical data analysis and first-scale prediction; edge computing nodes are deployed on the building cluster side to perform real-time data cleaning and second-scale prediction; and edge intelligent control devices are used to perform third-scale prediction. The duration of the first scale is longer than that of the second scale, and the duration of the second scale is longer than that of the third scale. Specifically, the cloud computing server stores massive amounts of historical data, runs a Long Short-Term Memory (LSTM) network model for day-ahead prediction (24-hour scale), establishes a digital twin model of photovoltaic-building cluster load, and performs spatiotemporal correlation analysis. Edge computing nodes are deployed on the building cluster side to perform data cleaning, anomaly detection, and cluster analysis; run a Transformer model for intraday prediction (4-hour scale), and perform rolling corrections based on Bayesian inference. The edge-side intelligent control device embeds a lightweight model that combines a convolutional neural network (CNN) and a gated recurrent unit (GRU) to achieve rapid response prediction on an ultra-short-term (15-minute scale).

[0024] Application Layer 4 provides users and maintenance personnel with visual interactive terminal devices and decision support services. It displays real-time dynamic sensing data and spatiotemporal correlation analysis views of photovoltaic output, total building load, individual building load, energy storage status, and environmental parameters through panoramic monitoring. It visualizes multi-level prediction results on both the supply and demand sides and the dynamic distribution of supply and demand matching indicators through supply and demand matching degree prediction results. It displays real-time prediction curves and measured curves of photovoltaic output and building load; and visualizes the dynamic changes of the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree. This mainly includes energy data visualization, i.e., real-time display of multi-source heterogeneous data and spatiotemporal correlation analysis views; supply and demand matching degree display, mainly including multi-level prediction result display and dynamic distribution display of supply and demand matching degree indicators; and dispatchable potential calculation and decision display, mainly including demand-side flexible dispatch potential curve display and optimized dispatch decision display.

[0025] In this embodiment, by adopting an integrated "source-load" design, the supply side (photovoltaics) and the demand side (building complex) are incorporated into a unified physical-information fusion framework, avoiding data silos and communication delays caused by independent systems. Furthermore, the devices at each level of this system work together to achieve closed-loop management from physical perception to optimized control.

[0026] Furthermore, based on the above embodiments, such as Figure 1 As shown, the control layer 3 in this embodiment adopts a cloud-edge-device collaborative computing architecture, wherein: the cloud computing platform is used to run long-term historical data analysis and first-scale prediction; the edge computing node device is deployed on the building group side to perform real-time data cleaning and second-scale prediction; the device-side intelligent control device is used to perform third-scale prediction; wherein the duration of the first scale is longer than the duration of the second scale, and the duration of the second scale is longer than the duration of the third scale.

[0027] Specifically, control layer 3 adopts a cloud-edge-device collaborative computing architecture. Based on the computing needs and data timeliness requirements of different prediction scales, computing tasks are distributed hierarchically to the cloud, edge, and device sides, forming a computing system of "long-term analysis - real-time processing - immediate response." The cloud computing platform stores massive amounts of historical data, runs LSTM models for day-ahead (first scale) prediction (24-hour scale), establishes a digital twin model of photovoltaic-building cluster load, and performs spatiotemporal correlation analysis. Edge computing nodes are deployed on the building cluster side, performing data cleaning, anomaly detection, and cluster analysis, running Transformer models for intraday (second scale) prediction (4-hour scale), and performing rolling corrections based on Bayesian inference. The device on the device side uses a lightweight model combining convolutional neural networks and gated recurrent units to achieve rapid response predictions for ultra-short-term (third scale) (15-minute scale). The duration of the first scale (day-ahead prediction) is greater than that of the second scale (intraday prediction) which is greater than that of the third scale (ultra-short-term prediction).

[0028] By adopting a cloud-edge-device layered computing architecture, the computing tasks can be reasonably allocated and coordinated, which not only ensures the stability of long-term predictions but also improves the response speed of real-time predictions. This avoids the limitations of a single computing architecture and provides solid computing power support for collaborative predictions across multiple time scales.

[0029] Furthermore, based on the above embodiments, such as Figure 1 As shown, in this embodiment, the sensing layer 1 includes: an irradiance meter, a sky imaging camera, a module temperature sensor, and an electrical measurement device deployed on the photovoltaic side, used to collect irradiance, cloud motion characteristics, module temperature, and output power; and smart meters, heating and cooling meters, environmental sensors, and personnel occupancy monitoring devices deployed on the building complex side, used to collect sub-item electrical load, heating and cooling load, indoor environmental parameters, and personnel occupancy status.

[0030] The sensing layer 1 includes photovoltaic sensing units, energy storage sensing units, load sensing units, environmental sensing units, and equipment sensing units. The photovoltaic sensing units collect data on irradiance, module temperature, output power, and inverter power. The energy storage sensing units collect data on state of charge (SOC), voltage, current, and temperature. The load sensing units collect data on electrical load, thermal load, sub-meters, and the number of personnel. The environmental sensing units collect data on room temperature, relative humidity, illuminance, and carbon dioxide concentration. The equipment sensing units collect data on coefficient of performance (COP), rotational speed, pressure, and start / stop status.

[0031] Specifically, based on the cyber-physical system (CPS) perception layer 1 physical sensing device, comprehensive status perception is achieved on both the photovoltaic side and the building complex side. On the photovoltaic side, a ground-based irradiance meter is deployed to collect parameters such as GHI, DNI, and DHI (sampling frequency 1 minute). A sky imaging camera captures cloud cover and cloud movement characteristics, and collects parameters such as ambient temperature, humidity, air pressure, wind speed, and wind direction (sampling frequency 1 minute). Simultaneously, smart sensors collect data on module temperature, string voltage / current, and inverter operating status. On the building complex side, smart meters, heating / cooling meters, and environmental sensors are used to collect real-time data on sub-item electrical load, heating / cooling load, indoor temperature and humidity, etc. Parameters such as concentration and personnel occupancy rate are collected. The aforementioned sensing data is uploaded in real time to edge computing node devices or cloud computing platforms via the CPS network layer 2 data transmission device.

[0032] To address the differences in photovoltaic module installation posture (different tilt angles / orientations / tracking types), the system introduces a geometric transformation algorithm in the preprocessing stage to uniformly convert the GHI collected by the sensing layer 1 into photovoltaic array planar irradiance (POA irradiance), and combines incident angle correction to eliminate data deviations caused by posture differences.

[0033] To address short-term fluctuations caused by rapid changes in cloud cover, sky imaging data and optical flow algorithms are used to analyze cloud motion characteristics and predict the cloud's shading effect on solar radiation in the coming minutes, providing forward-looking input for minute-level scheduling of CPS control layer 3. For example, in a scenario where a south-facing 15° tilted roof is used in conjunction with horizontal components, the correlation coefficient between the irradiance sequences of the two components increases from 0.62 to 0.85 after POA conversion (the closer the correlation coefficient is to 1, the stronger the correlation), significantly enhancing the physical consistency of multi-source data.

[0034] Furthermore, based on the above embodiments, such as Figure 1 As shown, in this embodiment, after the perception layer 1 completes data acquisition, the control layer 3 performs multi-source heterogeneous data processing 31, including data cleaning and correction, to achieve data alignment and deviation correction, then performs data feature clustering to generate composite cluster labels and a reference set of similar samples, and then performs data preprocessing to achieve outlier detection, missing value filling and multi-granularity sample generation.

[0035] Specifically, the collected data is cleaned at the edge of the CPS system control layer 3 or in the cloud. The NOCT (Nominal Operating Temperature) model is used to correct the efficiency degradation caused by temperature, as shown in formulas (1) and (2). The degree of component efficiency degradation when the actual temperature deviates from the reference temperature is quantified. The original output power signal is decomposed into irradiation driving term and temperature degradation term to improve the interpretability and accuracy of the model for changes in the thermal environment.

[0036] (1) (2) In the formula, This indicates the original output power of the photovoltaic module, in kW; The basic power generation capacity of a photovoltaic module is expressed in kW. This represents the power loss caused by the increase in component temperature, expressed in kW. This represents the corrected predicted photovoltaic power output, in kW; This represents the initial predicted output of photovoltaic power generation, in kW; k Indicates the temperature coefficient of a photovoltaic module; Indicates photovoltaic modules t The actual temperature at that moment, in °C; Indicates the nominal operating temperature.

[0037] By combining hourly weather forecast data with minute-level data from ground meteorological stations, and performing time-series alignment and bias correction on hourly weather forecast data, a composite feature of "long-term trend + near-term observation" is formed. This provides reliable input for multi-timescale forecasts, ensuring the stability of day-ahead forecasts while providing real-time correction for intraday or ultra-short-term forecasts. For example, when the meteorological observatory forecasts clear skies but ground stations observe localized cloud cover increases, the system increases the weight of cloud cover features within 10 minutes to reduce forecast bias caused by cloud cover.

[0038] Integrating data from the building, environment, and photovoltaic sides into a unified time base involves aligning the timestamps of multi-source data. For example, using "the 30th second of every minute" as the time node for data collection and statistics, a coupled "source-load-environment" data view is formed, providing support for spatiotemporal correlation analysis and thus improving the interpretability of the analysis results.

[0039] Leveraging the computing power of the CPS control layer 3 collaborative computing and control device architecture, K-means clustering analysis is performed on massive historical data to effectively extract similar features from photovoltaic power generation and building energy consumption data. This allows all datasets to be classified according to these similar features. By processing and analyzing data from different clusters, higher-quality data is provided for training the predictive model.

[0040] The collected datasets typically generate high-dimensional databases. When using traditional distance-based clustering methods, high-dimensional data can lead to increased computational load and decreased clustering performance. Therefore, techniques such as subsequence generation and dimensionality reduction are employed. Figure 2 and Figure 3 As shown, the statistical features (maximum / minimum / mean / standard deviation) of daily load data are extracted, transforming high-dimensional time-series data into a low-dimensional feature matrix, which significantly reduces the computational load.

[0041] Specifically, the daily load data, consisting of n points, is divided into α time periods, and all load data within each time period is aggregated into a single data point pi. Thus, the daily load data is integrated into a subsequence of α points, with each data point pi containing the data characteristics corresponding to that time period. Figure 2 A schematic diagram illustrating the subsequence generation process is provided. Multiple subsequences are generated for all load data using the same method. Subsequently, statistical characteristic values ​​such as maximum, minimum, mean, and standard deviation are extracted from each subsequence to characterize the daily load data. K-means clustering analysis is used to cluster these statistical characteristic values. Through multiple iterations, the optimal clustering result is reached when the cluster centers no longer change. The K value is selected using a combination of the elbow method and the silhouette coefficient.

[0042] In terms of clustering performance, it can identify data characteristics across different time periods, thereby improving clustering accuracy. In terms of clustering efficiency, it can effectively reduce the dimensionality of the load data. Figure 3 This demonstrates the specific process of dimensionality reduction for load data, for a dataset containing... m The dataset of daily load data, raw m × n 1-th order matrix D Divided by time α A time period is aggregated to generate a simplified version. m × α Order subsequence matrix D´ Further extraction from the subsequence matrix β A statistical feature can reduce the matrix dimension from m × α Down to m × β ( β < α < n ), to obtain the final m × β Clustering feature matrix D" Clustering analysis based on data feature matrices significantly improves clustering efficiency. Specifically: m The number of days in the dataset, i.e., the number of rows in the matrix, represents the total number of samples participating in the cluster analysis; n The number of sampling points or feature dimensions representing the original daily load data, i.e., the number of columns in the original matrix; x 11 This represents the data in the first row and first column of the matrix; x mn Represents the first in the matrix m Line number n The data in the column will not be described individually for the other elements; αThis represents the number of time periods into which the daily load data is divided (e.g., if a day is divided into 4 time periods, then...). α =4), which is the number of columns in the subsequence matrix after the first dimensionality reduction; sum 11 This represents the data in the first row and first column of the subsequence matrix; sum mα Represents the first subsequence in the subsequence matrix m Line number α The data in the column will not be described individually for the other elements; β This represents the number of statistical features extracted from each subsequence (e.g., maximum, minimum, mean, standard deviation). β =4), which is the number of columns in the final clustering feature matrix; feature 11 This represents the data in the first row and first column of the clustering feature matrix; feature mβ Represents the first cluster in the cluster feature matrix. m Line number β The data in the column will not be described individually for the other elements.

[0043] After completing data feature clustering, composite cluster labels are generated for different clusters of data, consisting of "sky conditions (sunny / cloudy / overcast) + human behavior patterns (weekday / weekend, lunch peak / evening peak) × equipment status (limited power, maintenance)". A "reference set of similar samples" is then established for each cluster in the CPS database, ensuring that similar samples use consistent preprocessing strategies and model weights in subsequent steps. This lays a template-based foundation for subsequent data preprocessing, defining interpolation methods, anomaly thresholds, and feature weights by cluster, avoiding distortion caused by mixing samples from cloudy or sunny scenarios with those from sunny days. Furthermore, the composite cluster labels are used as one of the features in the spatiotemporal correlation modeling steps on both the supply and demand sides, improving the interpretability and robustness of the spatiotemporal modeling.

[0044] The next step is data preprocessing, specifically outlier detection. After completing K-means clustering and obtaining composite cluster labels, a "similar sample dataset" is created for each cluster. A combined strategy of "outlier detection + state labeling + feature preservation" is used to process outlier data, avoiding misjudging normal short-term fluctuations in multi-cloud scenarios as anomalies. It also avoids mistaking equipment states such as maintenance or power limitations for environmental fluctuations, thus improving the robustness of subsequent prediction and matching calculations from the source.

[0045] First, outliers are screened within the cluster using similar days as a reference. For any time t, if the measured value y(t) satisfies equation (3), it is determined to be an outlier. This is achieved through the definition of "similar days within the cluster". With σ(t), the threshold is automatically adapted to differences in modes such as sunny / cloudy and weekday / weekend, avoiding threshold distortion caused by cross-mode mixed statistics.

[0046] (3) In the formula, y ( t )express t The measured value at that moment; Indicates similar days in the same cluster t The average value of the data at time σ ( t ) indicates similar days in the same cluster t Standard deviation of the data at any given time.

[0047] Secondly, a sliding window statistical approach is introduced to adapt to the rapid changes in minute-level data, avoiding the erroneous rejection of normal short-term drops or spikes caused by cloud shadows. Using a 5-minute sliding window, the local mean and standard deviation within the window are calculated, and combined with the first-order difference rate of change at the window edge, "sudden but interpretable" fluctuations are retained and labeled, thereby preserving high-frequency information that is crucial for ultra-short-term forecasting.

[0048] Next, the causes of anomaly candidate points are differentiated and labeled: environmental / system fluctuations (such as radiation spikes caused by cloud cover, and load increases due to a sudden increase in personnel) and equipment status anomalies (such as inverter maintenance shutdown, power outages due to component cleaning, inverter power limiting, and communication interruptions) are assigned different labels (e.g., "radiation spikes," "inverter power limiting," "equipment maintenance," etc.). For equipment status anomalies, unlike the traditional crude approach of "directly deleting outliers," the model retains the anomaly segment and its preceding and following context features, enabling the model to learn the complete pattern of "before the anomaly occurs - during the anomaly occurs - after the anomaly ends." This ensures that the model can recognize both normal scenarios and learn anomaly patterns in abnormal scenarios, thereby improving the robustness and stability of the model under complex operating conditions.

[0049] Missing value imputation. To address missing data within clusters, a combination of interpolation, in-cluster machine learning imputation, and physical constraints is employed to ensure that the imputed data not only conforms to the typical pattern of the cluster but also satisfies the physical boundaries of photovoltaic power generation and building load, providing consistent and contradictory data for subsequent prediction model training.

[0050] First, for short-term missing data (e.g., the number of consecutive missing data points does not exceed the threshold L), linear interpolation is used: the specific calculation formula is shown in (4): (4) In the formula, express t The measured value at time 1; express t The measured value at time 2; t 1 and t 2 represents missing points. t The adjacent valid times before and after the current time.

[0051] Secondly, for long missing segments or segments significantly affected by environmental or human factors, similar sub-sequence samples are selected within the same composite cluster label to construct a K-Nearest Neighbor (KNN) imputation model with inputs such as POA irradiance, module temperature, cloud cover, inverter status, occupancy rate, and setpoint. This model estimates the missing segments to ensure consistent scene representation. Physical constraints are also applied, such as ensuring non-negative photovoltaic power, that photovoltaic output approaches zero when nighttime POA irradiance is approximately zero, and that building power and heating / cooling loads meet equipment rated power and operating condition constraints. This avoids unreasonable imputation results such as "nighttime power generation" or "negative power."

[0052] Multi-granularity sample generation. After unifying timestamps, multi-granularity sample sets are generated from minute-level raw data: 1-minute granularity is used for 15-minute ultra-short-term prediction and edge control; 15-minute granularity is used for 4-hour intraday prediction, matching degree calculation, and rolling correction; 1-hour granularity is obtained by aggregating 15-minute data and is used for day-ahead prediction training and inference 24 hours in advance. When constructing training / validation datasets, composite cluster labels are used as the hierarchical basis, and the training set, validation set, and test set are divided according to a fixed proportion or fixed number of days extracted from each cluster to ensure consistent coverage of each cluster in different datasets. At the same time, the sampling probability of clusters with few samples but significant impact on the system (e.g., "cloudy + lunch peak + power limit / maintenance") is increased or given higher weight in the loss function, so that the model does not overfit common clusters in key scenarios, nor ignore low-probability but high-risk clusters; composite cluster labels, as part of the node / time features of the spatiotemporal correlation modeling graph model and the multi-timescale hierarchical prediction input, are used throughout subsequent modeling and calculation, thus forming a closed-loop consistent link of "clustering-preprocessing-modeling-computation".

[0053] Furthermore, based on the above embodiments, such as Figure 1 As shown, in this embodiment, when the control layer 3 executes the construction of a photovoltaic-building load coupling model that integrates time and space dimensions, it is specifically used to: perform cluster analysis on historical and real-time data to generate composite cluster labels containing weather conditions, personnel patterns, and equipment status; calculate the cross-correlation function of photovoltaic output sequence and building load sequence under different time delays based on composite cluster labels to obtain scenario-based time lag characteristics; construct an equivalent fusion distance characterizing the relationship between electrical connection and spatial propagation by combining distribution network topology, geographic information, and cloud motion spatial propagation characteristics; and dynamically learn and output the spatiotemporal correlation weight matrix between photovoltaic nodes and load nodes using a graph attention network, taking composite cluster labels, time lag characteristics, and equivalent fusion distance as inputs, as the photovoltaic-building load coupling model.

[0054] Specifically, a digital twin mapping of photovoltaic output and building load is constructed in the CPS control layer 3. The accurate characterization of the physical world to the digital space is achieved through spatiotemporal correlation modeling 32 on both the supply and demand sides. This includes analyzing the time correlation to identify time lag characteristics, modeling the spatial correlation to calculate transmission loss and equivalent fusion distance, and dynamically obtaining the dynamic correlation weight matrix of "source-load" through spatiotemporal correlation fusion.

[0055] The time correlation analysis includes: In the time dimension, a cross-correlation function between photovoltaic power output and building load is constructed, as shown in formula (5). Then, the linear correlation between photovoltaic power output and building load under different time delays is analyzed, the optimal lag time is identified, the degree of difference between supply and demand on both sides in the time dimension is calculated, and the correlation characteristics between photovoltaic power output and building load in daily, weekly, and seasonal cycles are analyzed.

[0056] (5) In the formula, This represents the cross-correlation function value between photovoltaic power output and building load power sequence; For time delay; N Indicates the total number of samples in the sequence; express t Photovoltaic output power at any given time, in kW; The average value of the photovoltaic power series is expressed in kW. express Building load power at any given time, in kW; This represents the average value of the load power sequence, in kW.

[0057] The larger the value (positive value), the greater the delay. The stronger the synergy between photovoltaic power output and building load (peak photovoltaic power output corresponds to peak load), the better. The smaller (negative) the value, the greater the delay. The stronger the complementarity between photovoltaic power output and building load (the peak value of photovoltaic power output corresponds to the valley value of the load); The closer to 0, the greater the delay. Under these conditions, there is no obvious linear relationship between photovoltaic output and building load. By calculating the cross-correlation function under different time delays, the optimal time matching point between photovoltaic and building load is determined, the degree of supply and demand coordination and complementarity is calculated, and the correlation pattern under different time scales (daily, weekly, and short-term) is further analyzed, providing data support for the overall scheduling of photovoltaic, energy storage, and building flexible loads.

[0058] Using composite cluster labels as the grouping basis, cross-correlation function values ​​are calculated within each cluster, and the optimal lag time of source load under different scenarios is obtained. This significantly distinguishes the differences in patterns such as "sunny / partly cloudy / overcast" and "weekday / weekend," thereby forming the time prior relationship required for "supply and demand matching degree prediction." That is, it clarifies the relative lag pattern of photovoltaic output and building load on the time axis under different scenarios. On the one hand, it helps the model learn the phase relationship of source load under different scenarios during training, reducing the prediction bias caused by "time misalignment." On the other hand, it provides an alignment reference when calculating the time matching degree, enabling the matching degree prediction to more accurately distinguish between "true supply and demand mismatch" and "supply and demand differences caused by different occurrence times."

[0059] Spatial correlation modeling includes: Based on the distribution network topology and geographic information obtained from CPS network layer 2, a spatial mapping relationship between "PV source points and building cluster load points" is constructed. The distribution network and transmission network topology are incorporated into the prediction model input. The connection between PV and building cluster loads is no longer described solely by geometric distance, but rather by electrical information such as power network transmission paths and total impedance. This electrical information is directly used as model input, allowing the model to more accurately depict the spatial propagation patterns of PV output fluctuations. Simultaneously, the spatial propagation characteristics of cloud motion and irradiance fields are introduced to express the spatial sequence of influence of regional cloud shadows on different PV source points. This enables the prediction model to accurately characterize the impact of cloud motion on PV output fluctuations when predicting supply-demand matching.

[0060] Introducing the transmission loss coefficient from the photovoltaic source point to the building cluster load point makes transmission loss a key factor in matching degree calculation, and its functional expression is shown in Equation (6). The transmission loss coefficient represents the power transfer from the photovoltaic node to the load point of the building cluster. i to load node j The proportion of losses during the process.

[0061] (6) In the formula, This represents the transmission loss coefficient from the photovoltaic source point to the building cluster load point, typically 0 < <1, the closer to 1, the smaller the loss; This represents the actual length of the power network transmission path. Indicates the voltage level transmitted in the power network; The ratio of the actual load to the rated capacity of a transmission line, expressed in kW; This represents the total impedance of the transmission line; the higher the impedance, the greater the loss.

[0062] In addition, the geographical distance and electrical distance are combined into an equivalent fused distance, so that the spatial weight can reflect the distance between nodes in the geographical location and the power network transmission distance between the source point and the load point. The specific formula is shown in (7).

[0063] (7) In the formula, Indicates the source node i to load node j The equivalent fusion distance between them is a unified calculation index that integrates geographical and electrical features; Indicates the source node i to load node j The geographical straight-line distance between them; Indicates the source node i to load node j Electrical distance between them; , Let represent the weighting coefficients of the two distances, and satisfy . .

[0064] Spatiotemporal correlation fusion includes: To achieve a unified expression of the temporal and spatial correlation between photovoltaic power generation output and building load, a graph attention network (GAT) is used to construct a source-load spatiotemporal correlation fusion model, which outputs dynamic source-load association weights that can directly serve the prediction of supply and demand matching, as follows: First, the graph nodes and graph structure are defined. Physical nodes (photovoltaic arrays, building loads) are mapped to graph nodes, and physical connections (distribution network lines) and logical relationships (geographic / cloud shadow propagation) are mapped to edges, so that the graph structure can simultaneously represent the electrical coupling relationship of the power grid and the spatial propagation law of local meteorological (cloud shadow) phenomena.

[0065] Secondly, feature construction. Node features and edge features are constructed as inputs for fusion modeling. Node features consist of source-load sequences, composite cluster labels, and daily-scale statistical features, used to characterize the dynamic state of photovoltaic and demand-side loads, as well as weather-human-equipment state scenarios. Edge features include network paths, total impedance, transmission loss, and equivalent fusion distance, used to ensure that the learning results of the "source-load" association weights simultaneously conform to network connectivity, transmission costs, and spatial propagation laws.

[0066] Secondly, correlation fusion and weight learning. Within the sliding time window, the correlation strength between source nodes and load nodes is adaptively learned using GAT, forming a dynamic "source-load" correlation weight matrix that is updated over time. Based on the composite cluster label, scene weight switching and adaptive adjustment are realized, thereby suppressing correlation distortion caused by cross-scene mixing and improving the robustness of the model under conditions such as sunny days, cloudy days, overcast days, and equipment power limitation or maintenance.

[0067] Finally, the results are output and applied. The output includes a dynamic correlation weight matrix for "source-load", a set of time-lag features obtained according to the scenario, and a set of edge weight parameters for spatial constraints. The dynamic correlation weight matrix is ​​used to generate spatially enhanced source-load coupling input features, providing the hierarchical prediction model with inputs that are more consistent with the actual impact range, thereby improving the stability and generalization ability of photovoltaic output prediction, building cluster load prediction, and supply-demand matching degree prediction. At the same time, the set of time-lag features and the set of edge weight parameters are used to provide weight basis for spatial matching degree and comprehensive matching degree calculation, so that the matching degree prediction can comprehensively reflect the "source-load" correlation, network connectivity characteristics, transmission loss impact, and cloud shadow propagation impact, thereby ensuring the complete link of "modeling-prediction-index calculation" and reducing systematic bias.

[0068] Furthermore, based on the above embodiments, such as Figure 1 As shown, in this embodiment, when the control layer 3 performs multi-timescale collaborative prediction, it is specifically used to: on the cloud computing platform, use a long short-term memory network to perform a first-scale prediction of the first prediction span and a first-time granularity; on the edge computing node device, use a Transformer encoder-decoder network to perform a second-scale prediction of the second prediction span and a second-time granularity, and perform residual correction based on the result of the first-scale prediction; on the edge intelligent control device, use a lightweight model combining a convolutional network and a gated recurrent unit to perform a third-scale prediction of the third prediction span and a third-time granularity; wherein, the first prediction span is greater than the second prediction span, the second prediction span is greater than the third prediction span, the first time granularity is greater than the second time granularity, and the second time granularity is greater than the third time granularity.

[0069] Specifically, relying on the powerful computing power and multi-level collaborative capabilities of the CPS system's collaborative computing and control device architecture, and addressing the engineering requirements of "accurate and stable day-ahead forecasts, rolling corrections for intraday forecasts, and responses to rapid changes in cloud shadows for ultra-short-term forecasts," a multi-timescale hierarchical forecasting system ("day-ahead forecasting layer - intraday forecasting layer - ultra-short-term forecasting layer") is constructed. This system uniformly outputs forecast results for photovoltaic power output and building cluster load at different time granularities and provides an alignment sequence for directly calculating spatiotemporal matching indicators for supply-demand matching degree forecasting. This three-level system uses the obtained preprocessed data and composite cluster labels as scenario conditions, and the obtained time lag characteristics and "source-load" dynamic correlation weights as coupling constraints. This ensures that each layer of forecasting is completed under "consistent scenario and spatiotemporal consistency" input expressions, thereby reducing systematic forecasting biases caused by scenario switching, time misalignment, and spatial mismatch.

[0070] (1) The first-scale prediction with the first forecast span and the first time granularity, i.e. the day-ahead prediction with a 24-hour scale and a 1-hour granularity.

[0071] The day-ahead forecasting layer utilizes the LSTM algorithm for 24-hour and 1-hour granular day-ahead forecasts. Deployed in a CPS cloud computing center, leveraging its massive historical data storage and batch processing capabilities, and considering the strong time-series characteristics of the research subjects—where previous load demand significantly impacts next load forecasts—an LSTM deep neural network with time-series memory is employed to construct the day-ahead forecasting model. The LSTM neural network is driven by both physical correction features and human factors, making the day-ahead layer a "steady-state prior," providing a more accurate forecasting baseline for lower-level rolling corrections, and offering a "full-day supply-demand framework" for matching degree forecasting.

[0072] The day-ahead forecasting model's input parameters include preprocessed historical 30-day hourly data (PV output, building load, POA irradiance, and module temperature) combined with hourly weather forecasts, calendar data, and weekday / holiday features. Simultaneously, extracted daily-scale statistical features (maximum / minimum / mean / standard deviation) and historical composite cluster labels are introduced as conditional inputs to allow the model to learn baseline patterns under different "day type / weather / equipment statuses." Furthermore, obtained scenario-level time lag features are incorporated to achieve source-load time-series alignment when constructing the input matrix, reducing fitting bias caused by "source-load phase differences under the same day type." Based on the output "source-load" dynamic correlation weights and spatial constraint features, the available supply of distributed PV is mapped to the load-side convergence point, forming a joint feature of "effective supply-effective demand," making the day-ahead layer output more closely match the relevant information required for subsequent matching degree predictions. Output parameters include PV output and building load forecast curves for the next 24 hours.

[0073] The LSTM layers employ fully connected projection to maintain dimensionality consistency, with a hidden state dimension of 256. Normalization parameters consistent with multi-source heterogeneous data processing are used for both input features and target quantities. Composite cluster labels utilize learnable embedding vectors concatenated with continuous features. A fully connected output layer maps the hidden states to a dual-channel output of "PV power output and building load," and the dual-channel output is inversely normalized according to the feature scale. To verify the model's effectiveness and evaluate its prediction performance, the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) are selected as evaluation metrics. Detailed parameter settings for this layer are shown in Table 1 below.

[0074] Table 1. Parameters of the Forecast Model

[0075] (2) The second-scale prediction with the second prediction span and the second time granularity, namely the intraday prediction with a 4-hour scale and a 15-minute granularity.

[0076] The intraday prediction layer utilizes a Transformer encoder-decoder to provide intraday predictions with granularity of 4 hours and 15 minutes in advance. Deployed on a CPS edge computing node, it leverages the low latency and local inference capabilities of the device. Considering that intraday prediction requires rapid updates based on recent observations, while simultaneously integrating previous day prediction priors and minute-level perturbation information, and given the nonlinear long- and short-term dependencies between the source and payload sides, a Transformer encoder-decoder network with parallel modeling and attention filtering capabilities is employed to construct the intraday prediction model. The previous day layer output is used as the residual correction baseline, enabling the intraday layer to enhance its responsiveness to short-term trend changes while maintaining stability.

[0077] The intraday forecasting model's input parameters include the previous day's forecast results and minute-level measured data from the most recent 4 hours (output, POA irradiance, load, indoor environment, and occupancy rate, etc.), as well as composite cluster labels and daily-scale statistical features. The composite cluster labels are used to identify the "current day type / weather / equipment status scenario," allowing the model to adaptively adjust its focus under different scenarios. Furthermore, combining "source-load dynamic correlation weights" and cloud shadow propagation sequence information, photovoltaic nodes strongly correlated with the current load node and their spatial propagation information are used as attention-guiding features, thereby capturing the impact of cloud shadow disturbances on effective demand-side supply earlier. Output parameters include photovoltaic output and building load forecast curves with granularity of 15 minutes for the next 4 hours, and support for edge nodes to be updated on a 15-minute cycle.

[0078] The Transformer network model consists of four stacked encoder layers and four stacked decoder layers. The hidden layers in the feedforward neural networks of each encoder / decoder layer have a dimension of 512. Pre-LN structure is used for layer normalization to enhance training stability. The GELU activation function is used for the feedforward network. Position encoding uses sin-cosine and is superimposed with temporal semantic embeddings such as "time period / weekday". Composite cluster labels are injected into the encoder and decoder as learnable embedding vectors as conditional vectors. To suppress short-term noise and overfitting, the dropout probability is set to 0.1, and residual connections are used. The AdamW optimizer is used during training with a linear warmup followed by a cosine decay learning rate. Detailed parameter settings for this layer are shown in Table 2 below.

[0079] Table 2 Intraday Forecasting Model Parameters

[0080] (3) The third-scale prediction with the third prediction span and the third time granularity, namely, the ultra-short-term prediction with a 15-minute scale and a 1-minute granularity: The ultra-short-term prediction layer utilizes the GRU algorithm for ultra-short-term predictions with granularities of 15 minutes and 1 minute in advance. Running on a CPS-side intelligent control device or lightweight intelligent gateway, it leverages the device's real-time perception of the field status. Considering that ultra-short-term prediction requires minute-level rapid updates at the edge and sensitivity to rapid changes in cloud cover, short-term personnel fluctuations, and equipment status switching, a lightweight one-dimensional convolutional network is used to extract high-frequency local features. This feature is then cascaded with a GRU to model short-term sequence dependencies. Online learning is combined to achieve adaptive updates of the model during operation, enabling the ultra-short-term layer output to serve as a minute-level look-ahead measure for matching degree prediction and edge control.

[0081] Input parameters include measured data from the most recent 15 minutes (output, POA irradiance, cloud cover / motion characteristics from sky camera, load, and key equipment status) and the current composite cluster label. The composite cluster label drives the adaptive selection of the online update strategy, improving sensitivity to cloud motion characteristics and limiting overfitting in the "multi-cloud / rapid cloud shadow" cluster, and enhancing the tracking ability of load behavior and setpoint changes in the "stable under clear skies" cluster. Simultaneously, a "source-load dynamic correlation weight" is introduced to generate effective supply characteristics actually related to the current load node at the edge, reducing the error amplification caused by inferring matching degree solely from single-point output. Output parameters include a 1-minute granularity sequence of photovoltaic output and building cluster load prediction for the next 15 minutes, and a minute-level aligned sequence for matching degree prediction calculation.

[0082] A hybrid neural network model combining CNN and GRU was constructed. The CNN network uses a 3-layer stacked one-dimensional convolutional layer (Conv1d×3 structure) to maintain gradient stability through residual connections, followed by a 1-layer GRU structure to achieve short-term dependency modeling. The output layer uses linear mapping to generate minute-level prediction sequences for photovoltaic power output and building load, respectively, and performs normalization and denormalization at the edge side in the same manner as above. An SGD optimizer is used with L2 regularization and gradient pruning to prevent parameter overfitting and suppress model drift. When anomaly / state markers indicate equipment status such as "inverter power limit / maintenance / communication interruption", the update step size is reduced; when marked as environmental or behavioral disturbances such as "rapid changes in cloud cover / sudden increase in personnel", the weights of relevant features are increased and updates are maintained to achieve state-interpretable adaptive learning. The detailed parameter settings of the model are shown in Table 3 below.

[0083] Table 3 Parameters of the Ultra-Short-Term Prediction Model

[0084] Furthermore, to address the problems of existing solutions exhibiting fragmented predictions across different time scales, lacking hierarchical fusion and rolling correction mechanisms based on measured data, and thus struggling to adapt to dynamically changing scenarios, based on the aforementioned embodiments, as follows: Figure 1 As shown, in this embodiment, after the control layer 3 performs multi-timescale collaborative prediction, it is also used to: introduce real-time observation data, and perform rolling correction on the results of the first-scale prediction, the second-scale prediction, and the third-scale prediction through a Bayesian update framework; optimize the results of the rolling correction through a weight fusion mechanism of time decay and error adaptation, and generate the final prediction sequence of photovoltaic output and building load.

[0085] Specifically, a Bayesian update framework is used to calibrate prediction results in real time and adapt to scene changes. Adaptive weight adjustment is used to dynamically weight and fuse multi-level prediction results, achieving rolling correction across multiple time scales, as follows: Bayesian Update Framework: Based on the obtained three-layer prediction results (pre-day, intra-day, and ultra-short-term), a multi-timescale rolling correction mechanism is constructed at the CPS edge computing layer. Leveraging its proximity to the data source, the prediction results can be continuously calibrated with the latest observations, maintaining output continuity even under scene switching (based on composite cluster labels) and rapid changes in cloud shadows. Specifically, at each rolling update, using the "previous layer / previous time prediction" as the prior and the "current observation / near-end observation convergence" as evidence, a Bayesian update is used to obtain the posterior prediction. This transforms the three-layer prediction from "offline trained predictions" into "online usable predictions" that are calibrated in real-time and adapt to scene changes.

[0086] Bayesian updates assume that the predicted values ​​follow a Gaussian distribution within the current time window, and represent the prior prediction as... The observation is expressed as And calculate the posterior distribution. The mean of the posterior distribution is updated using the Kalman algorithm, as shown in formulas (8) and (9): (8) (9) In the formula, This represents the mean of the prior prediction results. It is the prediction value of the previous time step or the previous level, obtained by statistically analyzing the error between the model at each level and the historical predictions on the validation set, and calibrated according to the composite cluster label for different scenarios. It represents the variance of the prior prediction, that is, the uncertainty of the prior prediction. z Indicates the current observation value; This represents the observation variance, i.e., the uncertainty of the measured data, which is jointly determined by the acquisition noise and the intensity of near-end disturbances, and is based on the nearest sliding window. RMSE Perform adaptive updates;K Indicates Kalman gain; This represents the mean of the posterior prediction results, and the final predicted value is obtained after correction. This represents the variance of the posterior prediction result, i.e., the uncertainty of the posterior prediction result.

[0087] Adaptive weight adjustment: In order to enable the three-level forecasts to form a final forecast result that can be directly used for matching degree calculation under the same time base, a weight fusion mechanism of "time decay + error adaptation" is further constructed to dynamically weight and fuse the day-ahead, intraday and ultra-short-term forecasts to generate the final rolling forecast sequence and realize the dynamic weighted fusion of multi-level forecast results.

[0088] Regarding the time interval between the predicted time and the current time Define time weights As shown in equation (10), the smaller the time interval, the closer the time. The larger the value, the longer the time interval, and the further back in time. The smaller the value.

[0089] (10) In the formula, This represents the time interval between the predicted time and the current time. λ Indicates the attenuation coefficient, calibrated by layer: ultra-short-term layer λ Small, slow decay, daytime layer λ Large, and decays quickly.

[0090] Meanwhile, based on the prediction errors of each layer within the most recent sliding window ( RMSE Define error weights As shown in equation (11), the smaller the error ( RMSE (smaller) The larger.

[0091] (11) In the formula, RMSE This represents the prediction error of the layer within the most recent sliding window (the smaller the value, the higher the accuracy). This represents the calibration coefficient, calibrated by the performance of the validation set, so that the layer with smaller error receives higher weight in the current scenario.

[0092] Overall weight The result, obtained after normalization according to equation (12), is used to fuse the three-layer prediction outputs at the same granularity. Specifically, the 1-hour granularity is based on the day-ahead layer with added intra-day bias correction; the 15-minute granularity is based on the intra-day layer with added near-end correction of the ultra-short-term layer; and the 1-minute granularity is based on the ultra-short-term layer and maintains consistency with the intra-day / day-ahead layer trends. After fusion, physical boundary constraints such as non-negativity and near-zeroing at night are applied to the photovoltaic output, thereby outputting the final source-load prediction sequence that can be directly used for spatiotemporal matching degree calculation.

[0093] (12) Furthermore, to address the lack of spatial matching quantification calculations in existing solutions and the absence of quantification indicators that consider geographical distance and transmission loss, based on the above embodiments, as follows: Figure 1 As shown, in this embodiment, when calculating the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree (i.e., supply and demand spatiotemporal matching degree), the control layer 3 is specifically used as follows: within the calculation window, after aligning the photovoltaic output and building load power sequences in the final prediction sequence in time, the normalized cross-correlation coefficient is calculated as the time matching degree; combined with the equivalent fusion distance, the photovoltaic supply and load demand in the final prediction sequence are weighted by spatial attenuation weights to calculate the matching ratio as the spatial matching degree; and the time matching degree and spatial matching degree are weighted and summed to obtain the comprehensive spatiotemporal matching degree index.

[0094] Specifically, in the analysis module of CPS control layer 3, the final rolling forecast sequence is used as the calculation object. Under a unified time reference, an interpretable matching degree calculation is performed on the "supply side (photovoltaic output) - demand side (building cluster load)". The output is a set of matching degree indicators that can be directly used for strategy discrimination and dispatchable potential calculation. The matching degree calculation uses composite cluster labels as scene indexes and time lag characteristics and source-load dynamic correlation weights as alignment and weighting criteria. This avoids misjudgments caused by static measurements based solely on power differences and ensures the engineering usability of the indicators under multiple scenarios and spatial constraints.

[0095] The calculation process for the time matching index is as follows: Time matching index TMI The Time Matching Index (TMI) is a core indicator for assessing the consistency of the time dimension between photovoltaic output (supply side) and building load (demand side). Through normalization calculations, it eliminates the influence of power magnitude, transforming the source-load time matching effect into an intuitive value within the [0,1] range. It eliminates the need to focus on power magnitude, allowing direct comparison of source-load time matching effects across different buildings and scenarios, providing a quantitative basis for subsequent strategy judgment and dispatchable potential calculation. Simultaneously, TMIUsing composite cluster labels as scene indexes, group calculations avoid global averaging masking local defects, accurately locating low-matching scenes. Furthermore, scene-level time lag features are combined to align the source load sequence, distinguishing between "time misalignment" and "true mismatch," avoiding misjudgment in static measurements, and ensuring the indicators remain practical in engineering applications under various scenarios such as weekdays / weekends, sunny / cloudy days, and equipment power limitations / maintenance. See formula (13): (13) In the formula, T represents the total number of moments within the calculation time window.

[0096] in, and Taken from the same granularity prediction sequence, the calculation period is of length [length missing]. T The calculation window, TMI The closer the value is to 1, the more synchronized the supply-side output and demand-side load are in time within the window. The easier it is for the system to achieve local consumption and peak shaving through means such as energy storage charging and discharging, load shifting and reduction. Conversely, it indicates that there is a significant time mismatch, requiring more flexible resources to participate or cross-time adjustment.

[0097] The calculation process for the spatial matching degree index is as follows: Spatial matching degree index SMI The Spatial Matching Index (SMI) is a core indicator for assessing the spatial coupling consistency between photovoltaic output and building load in a spatial dimension. The specific formula is shown in equation (14). It incorporates the impact of geographical distance, electrical path, and transmission loss on effective supply, thus distinguishing between situations where "supply appears sufficient but is spatially inaccessible / too costly" and "feasible spatial matching," providing a quantitative basis for local consumption and neighboring dispatch strategies on the distribution network side. Meanwhile, SMI By introducing equivalent fusion distance and source-load dynamic association weights, the spatial terms are dynamically adjusted as cloud shadows propagate and scenes switch, avoiding distorted calculations caused by using only fixed distance weights, and ensuring that the spatial matching degree can reflect the true reachability in the prediction stage.

[0098] (14) In the formula, Indicates the predicted photovoltaic source point i The output power, kW; Indicates the predicted building load point j The demand, kW; Indicates photovoltaic source point i to load point j The equivalent fusion distance; The spatial scale parameter is determined by the distribution network topology and transmission loss. The smaller the distance, the more significant the decrease in weight.

[0099] The core component of the spatial matching index is the exponentially decaying spatial weight. This indicates that the longer the path, the lower the matching effectiveness. Spatial distance is converted into an "effectiveness weight" through exponential decay, thus normalizing the spatial matching degree (value range [0,1]). This indicator does not simply look at "whether the total photovoltaic output of a certain area equals the total load demand," but rather at "whether the photovoltaic output can be transmitted to the load point in a low-loss, accessible manner." The higher the matching degree (stronger accessibility, lower loss), the better. SMI The closer to 1, the lower the fit (long distance, high loss, unreachable path), and the closer SMI is to 0.

[0100] The calculation process for the comprehensive matching index is as follows: Comprehensive matching index CMI (Comprehensive MatchingIndex) is the core indicator for "evaluating the degree of comprehensive spatiotemporal matching under the joint constraints of temporal matching and spatial reachability". The specific formula is shown in Equation (15). It expresses the rolling prediction sequence (reflecting future supply and demand relationship) and spatiotemporal weight information (reflecting spatial dynamic coupling) in the same evaluation index, and transforms the multi-dimensional matching effect into a single indicator that can be directly used for scheduling optimization, avoiding decision-making difficulties caused by conflicts between multi-dimensional indicators. CMI When the levels are too low, it can be determined that "there is a significant spatiotemporal mismatch between supply and demand within the future window," requiring compensation for this mismatch through energy storage charging and discharging and demand-side adjustment strategies. CMI When the value is too high, it can be determined that "the system has a high self-matching capability," and the strategy can prioritize low-cost, low-disturbance adjustment methods. The output results are directly linked to the scheduling strategy to achieve a closed loop of "computation-decision-execution."

[0101] (15) In the formula, and They represent TMI and SMI The weighting of the indicators, the allocation of their importance in the comprehensive calculation, and + =1, guaranteeing CMI Ultimately, it still falls within the [0,1] interval, and its value is determined comprehensively based on historical data and optimization objectives under different scenarios. CMI It can accurately reflect the actual matching capability in different scenarios, avoiding the distortion of "one-size-fits-all" calculation.

[0102] Furthermore, to address the shortcomings of existing solutions in assessing schedulable potential and the lack of precise mining and quantitative calculation of flexible load resources, based on the above embodiments, such as... Figure 1As shown, in this embodiment, control layer 3 is also used to: call the physical models of building thermal storage, energy storage equipment, and adjustable loads, and generate a demand-side flexible adjustment potential curve by combining real-time equipment status and indoor comfort constraints; generate a corresponding scheduling strategy based on the comprehensive spatiotemporal matching degree and the flexible adjustment potential curve; correspondingly, application layer 4 is also used to visualize the flexible adjustment potential curve and scheduling strategy. Specifically, when generating the demand-side flexible adjustment potential curve, control layer 3 is used to: multiply the theoretical adjustable power, comfort sensitivity coefficient, and real-time availability coefficient for each type of flexible load resource to obtain the effective adjustable power of the corresponding flexible load resource; and aggregate and optimize the effective adjustable power of the flexible load resource by combining the time-period electricity price signal, comprehensive spatiotemporal matching degree, and equipment operation constraints to generate the demand-side flexible adjustment potential curve.

[0103] Specifically, the dispatchable potential calculation 36 includes flexible load dispatch capacity calculation and flexible adjustment potential curve generation, which respectively realize the quantification of adjustment capacity, calculation of flexible load index FLI, price elasticity model, and indoor comfort constraints.

[0104] The calculation of the flexible load regulation capacity of the building complex includes: Based on the sub-metering data and equipment operating condition information collected by the physical sensing devices in the CPS sensing layer 1, and utilizing the digital twin model of the CPS control layer 3, a flexible resource library on the demand side is first established. This library is then refined and identified according to its response characteristics. These resources include building envelopes and interior furniture with thermal inertia (passive energy storage), active cold / heat storage tanks and electrochemical energy storage (active energy storage), electric vehicles and washing machines that can be time-shifted (transfer type), non-critical lighting and office equipment that can be interrupted for short periods (reduction type), and HVAC systems whose power can be adjusted via frequency conversion (regulation type). These various resources can provide load frequency regulation, load reduction, and load transfer at different time scales, ranging from seconds to hours.

[0105] Secondly, differentiated adjustment capacity calculation models are constructed for different types of flexible resources. For building thermal storage bodies, a second-order RC thermal network model is used to identify equivalent heat capacity and thermal resistance. Under the constraint of human thermal comfort (PMV index or room temperature fluctuation ±1.0℃), its charging and discharging heat potential as "virtual energy storage" is calculated. For active energy storage devices, the remaining available adjustment capacity is calculated based on real-time state of charge (SOC) and charging and discharging efficiency. For interruptible and relocatable electrical equipment, its adjustable power range is calculated by combining rated power, current load rate, and user-preset minimum / maximum operating time constraints.

[0106] Based on this, in order to achieve unified scheduling decisions for heterogeneous flexible resources, the discrete adjustable capabilities of each device are mapped to a unified dimensional system, thus constructing a flexible load index. FLI(Flexible Load Index), the specific calculation formula is shown in Equation (16). This index transforms the theoretical physical capacity into a degree of flexibility that reflects "actual availability at the current moment and in the current scenario" through weighted normalization. Comfort weight is introduced to characterize the user's sensitivity to adjustment, and availability coefficient is introduced to characterize the probability that the device is occupied or controlled. FLI The calculation parameters are dynamically updated with the composite cluster labels (such as weekdays / holidays, people in / out of the room), ensuring that the calculation results can truly reflect the source-load mismatch compensation capability of the demand side in a specific scenario.

[0107] (16) In the formula, Indicates the first k The theoretical adjustable power threshold for flexible load resources, kW; Indicates the first k The comfort weighting coefficient for flexible load-bearing resources is lower the more sensitive the resource is to comfort. Indicates the first k The availability coefficient of flexible load resources reflects whether the flexible load resources are currently occupied and whether they are subject to policy constraints. If they are not occupied or have no constraints, the coefficient is 1. This represents the peak building load within the calculation window, in kW.

[0108] The generation of the demand-side flexible adjustment potential curve includes: Leveraging the command generation and distribution capabilities of CPS control layer 3, within adjustable boundaries and FLI Based on the index, a price elasticity model is established and indoor comfort constraints are introduced (such as allowable deviation of air conditioning setpoints by ±2℃, lower limit of lighting illuminance, etc.). Combined with time-period electricity / carbon price signals and external incentives such as occupancy rate, the range of power that can be adjusted up / down and the maximum duration of demand side in each time period are output, forming a building-side flexible adjustment potential curve.

[0109] Furthermore, the potential curve is compared with the overall matching degree. CMI The system deeply couples the system with distribution network constraints (such as transformer capacity and line current carrying capacity) to generate executable optimized scheduling strategies and priority sequences, achieving a closed loop of "first calculating the degree of mismatch, then matching adjustable resources". This ensures that the scheduling strategy can effectively alleviate source-load mismatch while minimizing the impact on the user side.

[0110] Specific application examples of the system using the present invention are as follows: 1. (Mismatch between peak and low load of photovoltaic power output): On a sunny weekend at noon, the forecast results show that the photovoltaic power output reaches its peak, but the load of office buildings is at its low point. CMI(Extremely low), the system generates a "source-side consumption priority" strategy. First, it starts the building's thermal storage to "pre-cool / pre-heat" to store excess photovoltaic power (virtual energy storage). Second, it guides electric vehicle charging piles into "strong charging mode". Finally, it calls on mobile loads such as washing machines to maximize the local consumption of photovoltaic power.

[0111] 2 (Plunge in PV output combined with peak load): On a cloudy or rainy weekday afternoon, forecasts indicate a sudden drop in PV output while air conditioning load remains high. CMI (At lower levels), the system generates a "peak shaving and valley filling priority" strategy. First, it releases the electricity of the active energy storage device to fill the gap. Second, it adjusts the set temperature of the HVAC system (such as increasing it by 2°C) to reduce peak load. If necessary, it performs short-term interruptions on non-critical lighting and office equipment to ensure the safe operation of the distribution network.

[0112] 3. Peak Electricity Prices and Supply-Demand Balance Fluctuations: During peak electricity price periods on weekdays, fluctuations in photovoltaic power generation and load lead to slight mismatches. CMI (Medium) The system generates an "economically optimal" strategy, prioritizing the use of electrochemical energy storage for discharge arbitrage, while adjusting the variable frequency air conditioner to maintain power balance, avoiding frequent start-ups and shutdowns of high-power equipment, and achieving the lowest operating cost while ensuring comfort.

[0113] The system using the present invention has the following beneficial effects: 1. Improved prediction accuracy through multi-dimensional perception: By introducing multi-source heterogeneous features such as cloud motion, equipment status, and composite cluster labels, and combining graph attention network (GAT) to deeply model the spatiotemporal correlation of source load, and adopting a multi-timescale cascaded prediction architecture of "day-to-day-ultra-short-term", the problem that traditional single-point and single-dimensional prediction cannot cope with the randomness of photovoltaic fluctuations and building cluster loads is effectively solved, and the accuracy of source load prediction in complex scenarios is significantly improved.

[0114] 2. Interpretable Supply-Demand Matching Calculation: A novel dual-dimensional calculation index of temporal and spatial matching degree is proposed, and a comprehensive matching degree evaluation system is constructed. This system can not only intuitively quantify the temporal synchronicity of source and load, but also explicitly characterize spatial transmission accessibility and loss costs. It provides a clear and quantifiable decision-making basis for formulating strategies for source-load complementarity, local consumption, and cross-regional dispatching, avoiding the blindness of traditional judgments based solely on experience or single power difference values.

[0115] 3. Dynamic adaptive correction mechanism: A rolling correction framework based on Bayesian theory was established. Real-time observation data was used to update the multi-scale prediction results online. Combined with a scene-adaptive weight fusion mechanism, the prediction results were dynamically calibrated to mitigate disturbances such as sudden weather changes and equipment failures. This effectively suppressed the cumulative effect of prediction errors and ensured the stability and robustness of the system under all-weather conditions.

[0116] 4. Refined Adjustable Potential Exploration: A method for calculating the building's flexible resource adjustment potential is proposed, integrating physical mechanisms (RC thermal networks) with data-driven approaches. By constructing a flexible load index, heterogeneous building thermal storage, active energy storage, and various flexible devices are unified into standardized dispatchable resources. This not only achieves precise quantification of demand-side adjustment potential but also generates flexible adjustment potential curves that include power boundaries and duration. This enables the generation of optimal dispatch strategies that balance user comfort and grid economics, realizing the transformation of the building complex's integrated energy system from "passive energy consumption" to "active energy generation and consumption."

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photoelectric-building complex load spatiotemporal matching degree prediction system, characterized in that, include: The sensing layer is used to collect photovoltaic power output data and building cluster load data. The network layer is used to transmit the photovoltaic power output data and the building cluster load data to the control layer; The control layer is used to construct a photovoltaic-building load coupling model that integrates time and space dimensions based on the photovoltaic power output data and the building load data, perform multi-time scale collaborative prediction, obtain the prediction results of photovoltaic power output and building load, and calculate the time matching degree, spatial matching degree and comprehensive spatiotemporal matching degree of photovoltaic power output and building load based on the prediction results. When the control layer executes the construction of a photovoltaic-building load coupling model that integrates time and space dimensions, it is specifically used to: perform cluster analysis on historical data and real-time data to generate composite cluster labels that include weather conditions, personnel patterns and equipment status; and calculate the cross-correlation function of photovoltaic output sequence and building load sequence under different time delays based on the composite cluster labels to obtain scenario-based time lag characteristics. Using the composite cluster labels as conditions, the cross-correlation function of the photovoltaic power output sequence and the building load sequence under different time delays is calculated to obtain the scenario-based time lag characteristics, including: using the composite cluster labels as the grouping basis, calculating the cross-correlation function value in each cluster to obtain the optimal lag time of the source load under different scenarios, and clarifying the relative lag law of photovoltaic power output and building load on the time axis under different scenarios. Combining distribution network topology, geographic information, and cloud motion spatial propagation characteristics, an equivalent fusion distance characterizing the relationship between electrical connections and spatial propagation is constructed. Using a graph attention network, the composite cluster label, the time lag feature, and the equivalent fusion distance are taken as inputs to dynamically learn and output the spatiotemporal correlation weight matrix between photovoltaic nodes and load nodes, which serves as a photovoltaic-building load coupling model. The method of using a graph attention network to dynamically learn and output the spatiotemporal association weight matrix between photovoltaic nodes and load nodes, taking the composite cluster label, the time lag feature, and the equivalent fusion distance as inputs, includes: mapping physical nodes to graph nodes and mapping physical connections and logical relationships to edges; constructing node features and edge features as inputs for fusion modeling, wherein the node features consist of source-load sequences, composite cluster labels, and diurnal statistical features, and the edge features include network paths, total impedance, transmission loss, and equivalent fusion distance; and adaptively learning the association strength between source nodes and load nodes using a graph attention network within a sliding time window to form a spatiotemporal association weight matrix. When performing multi-timescale collaborative prediction, the control layer constructs a three-level prediction system: day-day, intraday, and ultra-short-term. The three-level prediction system includes a day-day prediction model, an intraday prediction model, and an ultra-short-term prediction model. The three-level prediction system uses the obtained preprocessed data and composite cluster labels as scenario conditions, and the obtained time lag features and source-load dynamic correlation weights as coupling constraints, and completes the prediction under a consistent scenario and spatiotemporal input expression. The inputs to the day-ahead forecasting model include: hourly data, hourly weather forecasts, calendar and weekday data, holiday features, daily-scale statistical features, composite cluster labels, time lag features, source-load dynamic correlation weights, and spatial constraint features. The output is a granular forecast curve of photovoltaic power output and building load for each hour of the next 24 hours. The hourly data includes photovoltaic power output, building load, POA irradiance, and module temperature. The inputs to the intraday forecasting model include day-ahead forecast results, minute-level data, cloud motion features, composite cluster labels, daily-scale statistical features, and source-load dynamic correlation. The system takes weights and cloud propagation information as inputs, outputting PV output and building load prediction curves at 15-minute granularity for the next 4 hours. The minute-level data includes PV output, POA irradiance, building cluster load, indoor environment, and occupancy rate. The ultra-short-term prediction model takes into account the measured data of the most recent 15 minutes, composite cluster labels, and source-load dynamic correlation weights as inputs, and outputs a PV output and building cluster load prediction sequence at 1-minute granularity for the next 15 minutes. The measured data of the most recent 15 minutes includes PV output, POA irradiance, cloud movement characteristics, building cluster load, and key equipment status. The application layer is used to visualize the prediction results, the temporal matching degree, the spatial matching degree, and the comprehensive spatiotemporal matching degree.

2. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 1, characterized in that, The control layer adopts a cloud-edge-device collaborative computing architecture, wherein: A cloud computing platform for running long-term historical data analysis and first-scale prediction; Edge computing nodes are deployed on the side of the building complex to perform real-time data cleaning and second-scale prediction; An end-side intelligent control device is used to perform third-scale prediction; Wherein, the duration of the first scale is greater than the duration of the second scale, and the duration of the second scale is greater than the duration of the third scale.

3. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 1, characterized in that, When performing multi-timescale collaborative prediction, the control layer is specifically used for: On a cloud computing platform, long short-term memory networks are used to perform first-scale predictions with the first prediction span and the first time granularity. In the edge computing node device, a Transformer encoder-decoder network is used to perform a second-scale prediction with a second prediction span and a second time granularity, and residual correction is performed with the result of the first-scale prediction as a baseline. On the edge intelligent control device, a lightweight model combining convolutional networks and gated recurrent units is used to perform third-scale predictions of the third prediction span and third time granularity. Wherein, the first prediction span is greater than the second prediction span, the second prediction span is greater than the third prediction span, the first time granularity is greater than the second time granularity, and the second time granularity is greater than the third time granularity.

4. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 2, characterized in that, After performing multi-timescale collaborative prediction, the control layer is also used for: Real-time observation data is introduced, and the results of the first-scale prediction, the second-scale prediction, and the third-scale prediction are rolled over and corrected using a Bayesian update framework. The rolling correction results are optimized by a weight fusion mechanism that combines time decay and error adaptation to generate the final predicted sequence of photovoltaic output and building load.

5. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 4, characterized in that, The control layer is specifically used for calculating the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree: Within the calculation window, after time alignment of the photovoltaic output and building load power sequences in the final predicted sequence, the normalized cross-correlation coefficient is calculated as the time matching degree. Based on the equivalent fusion distance, the photovoltaic supply and load demand in the final predicted sequence are weighted by spatial attenuation weights to calculate the matching ratio, which is used as the spatial matching degree. The temporal matching degree and the spatial matching degree are weighted and summed to obtain a comprehensive spatiotemporal matching degree index.

6. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 1, characterized in that, The control layer is also used for: By calling upon the physical models of building thermal storage bodies, energy storage devices, and adjustable loads, and combining them with real-time equipment status and indoor comfort constraints, a demand-side flexible adjustment potential curve is generated. Based on the comprehensive spatiotemporal matching degree and the flexible adjustment potential curve, a corresponding scheduling strategy is generated; Accordingly, the application layer is also used to visualize the flexible adjustment potential curve and the scheduling strategy.

7. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to claim 6, characterized in that, When generating the demand-side flexible adjustment potential curve, the control layer is specifically used for: For each type of flexible load resource, the theoretical adjustable power, comfort sensitivity coefficient, and real-time availability coefficient are multiplied to obtain the effective adjustable power of the corresponding flexible load resource. The effective adjustable power of the flexible load resources is aggregated and optimized by combining the time-period electricity price signal, the comprehensive spatiotemporal matching degree, and equipment operation constraints to generate a demand-side flexible adjustment potential curve.

8. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to any one of claims 1-7, characterized in that, The sensing layer includes: Irradiance meters, sky imaging cameras, module temperature sensors, and electrical measurement devices deployed on the photovoltaic side are used to collect irradiance, cloud motion characteristics, module temperature, and output power. Smart meters, heating and cooling meters, environmental sensors, and occupancy monitoring devices deployed on the side of the building complex are used to collect data on sub-items of electrical load, heating and cooling load, indoor environmental parameters, and occupancy status.

9. The photoelectric-building cluster load spatiotemporal matching degree prediction system according to any one of claims 1-7, characterized in that, The application layer is specifically used for: Real-time display of predicted and measured curves for photovoltaic power output and building load; The dynamic changes of the time matching degree, spatial matching degree, and comprehensive spatiotemporal matching degree are visualized.

Images