Image attention fusion and continuous learning cross-scenario workload prediction method and system

Abstract

The application discloses a cross-scene load prediction method and system based on image attention fusion and continuous learning, relates to the technical field of energy management, and comprises the following steps: obtaining operation feedback data of a target scene, generating an operation state feature vector and calculating a credibility factor; constructing a target scene image comprising a static attribute subvector and a behavior mode subvector; calculating the matching degree score of the target scene image and each source scene image in a preset model library, and selecting or integrating a target load prediction model; dynamically fusing the target scene image and the operation state feature vector based on an attention mechanism to obtain a fused feature vector; inputting the fused feature vector into the target load prediction model for prediction to obtain a load prediction result; and performing online incremental learning and parameter optimization on the target load prediction model. The application alleviates the technical problems of model matching inaccuracy, large deviation of the model under dynamic operation conditions and lack of self-adaptive optimization capability in the prior art.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a cross-scenario load prediction method and system based on profile attention fusion and continuous learning. Background Technology

[0002] In the refined scheduling and optimized operation of integrated energy systems in commercial buildings, accurate prediction of cooling and heating loads is crucial for achieving energy conservation and consumption reduction, and ensuring the stable and economical operation of the system. However, in practical engineering, quickly building a high-precision load prediction model for new buildings or existing buildings (target scenarios) without fully deployed monitoring systems often faces the challenge of "data scarcity" at the start. To address this issue, the current industry-standard approach is transfer learning, which involves attempting to apply mature prediction models trained on other similar buildings (source scenarios) to the target scenario after certain adjustments. While this method can shorten the model development cycle, it exposes a series of key technical bottlenecks during actual cross-scenario transfer, often resulting in prediction accuracy falling short of expectations.

[0003] First, existing technologies oversimplify and simplistically characterize "scene differences," leading to inaccurate model matching. Most current migration solutions rely on building type (e.g., office, shopping mall, hotel) label matching or curve similarity calculations based on limited historical load data. However, building load is a complex result of the interplay between physical characteristics and user behavior. Two buildings of the same type may exhibit completely different basic load characteristics due to differences in static physical properties such as building envelope insulation performance, window-to-wall ratio, and equipment system efficiency; they may also exhibit completely different load time-series patterns due to differences in user density variations and equipment start-up / shutdown strategies. Matching methods relying solely on single-dimensional labels cannot fully quantify these multi-dimensional and heterogeneous scene differences, resulting in "similar" source scene models that are not truly suitable, leading to a systematic bias in the migrated model from the outset.

[0004] Secondly, existing predictive models are mostly static "black boxes" trained on historical data, severely disconnected from the real-time operating status of energy systems. Commercial building energy systems do not operate in a fixed mode. Property managers dynamically adjust indoor temperature setpoints, modify equipment operating schedules, and even perform emergency start-ups and shutdowns based on factors such as weather, holidays, and actual occupancy. These real-time adjustments immediately impact the load, but traditional models trained on historical data typically only include macroscopic variables such as outdoor weather and date type as inputs, failing to perceive and respond to these microscopic, instantaneous human interventions or system state changes. This leads to significant discrepancies between predicted and actual values ​​under dynamic operating conditions, especially during periods of frequent scheduling, limiting their application value in real-time optimization control.

[0005] Furthermore, the migrated model lacks continuous adaptive and optimization capabilities, making it difficult to maintain long-term performance. A building's usage patterns and external environment are not static; they may experience slow seasonal changes, adjustments to operational strategies, or sudden shifts in patterns. Currently, the common practice is to fix the parameters after model migration and deployment, or to only perform costly "batch retraining" after accumulating a large amount of new data, which is impractical in real-world operations. Therefore, the model cannot evolve with changes in the target scenario itself, and its predictive performance gradually declines over time until it can no longer meet application requirements, necessitating another time-consuming model rebuilding process.

[0006] In summary, current cross-scenario load prediction technologies have significant shortcomings in all three key aspects of "matching, perception, and evolution". Summary of the Invention

[0007] To address the aforementioned technical problems in existing technologies, embodiments of the present invention provide a cross-scene load prediction method and system based on profile attention fusion and continuous learning. The technical solution is as follows: On one hand, a cross-scene load prediction method based on profile attention fusion and continuous learning is provided. The method includes: acquiring operational feedback data of a target scene and performing preprocessing and feature extraction to generate an operational state feature vector, and calculating a credibility factor for the operational state feature vector; constructing static attribute sub-vectors and behavioral pattern sub-vectors based on the inherent physical characteristics and behavioral patterns of the target scene, and constructing a target scene profile based on the static attribute sub-vectors and the behavioral pattern sub-vectors; calculating the matching score between the target scene profile and each source scene profile in a preset model library based on the credibility factor and a hybrid similarity matching algorithm of static similarity and morphological similarity, and selecting or integrating a target load prediction model based on the credibility factor and the matching score; the preset model... The library includes different source scenario profiles and corresponding pre-trained load prediction models. The target load prediction model is one load prediction model or an ensemble of multiple load prediction models from the preset model library. The target scenario profile is used as a key-value pair, and the running state feature vector is used as a query. Dynamic fusion based on an attention mechanism is performed to obtain a fused feature vector. The attention mechanism is used to calculate the matching degree between the query and the key, and the corresponding values ​​are weighted and summed based on attention weights according to the matching degree. The fused feature vector is input into the target load prediction model to predict the hot and cold loads of the target scenario, obtaining the load prediction result. Based on the load prediction result and multi-anchor point online elastic weight consolidation, online incremental learning and parameter optimization are performed on the target load prediction model.

[0008] Optionally, acquiring operational feedback data of the target scenario and performing preprocessing and feature extraction to generate an operational status feature vector, and calculating the credibility factor of the operational status feature vector, includes: acquiring operational feedback data of the target scenario; the operational feedback data includes environmental parameters, system status, control setpoints, and real-time load; performing data cleaning and multi-source time series alignment on the operational feedback data to obtain preprocessed data; performing feature extraction on the preprocessed data to generate an operational status feature vector; the operational status feature vector includes: environmental features, system status features, control features, real-time load features, and time-series derived features; and calculating the credibility factor based on the number of valid data points and the proportion of outliers in the operational status feature vector.

[0009] Optionally, the method further includes: adaptively adjusting the dimension of the running state feature vector based on the credibility factor; including: if the credibility factor is not less than a first preset threshold, then retaining the complete dimension of the running state feature vector; if the credibility factor is less than the first preset threshold and greater than or equal to a second preset threshold, then performing dimensionality reduction processing on the running state feature vector; if the credibility factor is less than the second preset threshold, then issuing a data quality alarm signal.

[0010] Optionally, static attribute sub-vectors and behavioral pattern sub-vectors are constructed based on the inherent physical characteristics and behavioral patterns of the target scenario, including: constructing static attribute sub-vectors based on the basic information of the target scenario, the thermal performance of the building envelope, the configuration parameters of the cooling and heating system, and the control and metering methods; the thermal performance of the building envelope includes the building thermal inertia index; constructing behavioral pattern sub-vectors based on the load and energy consumption time series patterns, personnel activity and space usage patterns, equipment system operation strategies, and environmental response patterns; the environmental response patterns include energy consumption response patterns differentiated based on the load-temperature sensitivity coefficient.

[0011] Optionally, based on the credibility factor and a hybrid similarity matching algorithm combining static similarity and morphological similarity, the matching score between the target scene image and each source scene image in the preset model library is calculated, including: calculating the static similarity between the corresponding static attribute sub-vectors of the target scene image and each source scene image in the preset model library; calculating the morphological similarity between the corresponding behavioral pattern sub-vectors of the target scene image and each source scene image based on the dynamic time warping distance; determining the matching weight based on the credibility factor; and performing a weighted fusion of the static similarity and the morphological similarity based on the matching weight to obtain the matching score between the target scene image and each source scene image in the preset model library.

[0012] Optionally, selecting or integrating a target load prediction model based on the credibility factor and the matching score includes: filtering multiple candidate models in the preset model library based on the matching score and a preset matching threshold; calculating a multi-dimensional score for each candidate model based on its matching score, historical performance score, and model robustness score; if the credibility factor is not less than a third preset threshold, then determining the candidate model with the highest multi-dimensional score as the target load prediction model; if the credibility factor is less than the third preset threshold but greater than or equal to a fourth preset threshold, then ranking the multi-dimensional score among the top N... top The candidate models are determined as base models, and the corresponding ensemble weights are calculated based on the multi-dimensional scores of each base model. Multiple base models are then integrated based on these ensemble weights to obtain the target load prediction model; N top It is an integer greater than 1.

[0013] Optionally, the target scene profile is used as a key-value pair, and the running state feature vector is used as a query. Dynamic fusion based on an attention mechanism is performed to obtain a fused feature vector, including: mapping the target scene profile, the historical running feature sequence of the target scene, and the running state feature vector to a unified dimension through a projection matrix to form a feature matrix; generating a query vector based on the running state feature vector, generating a key vector and a value vector based on the feature matrix, calculating the matching degree between the query vector and each key vector, and calculating the corresponding attention weight; performing a weighted summation of the value vector based on the attention weight to obtain a fused value vector; and concatenating the fused value vector with the running state feature vector to obtain the fused feature vector.

[0014] Optionally, based on the load forecasting results and online elastic weight consolidation of multiple anchor points, online incremental learning and parameter optimization are performed on the target load forecasting model, including: monitoring the average absolute error of the load forecasting results within a sliding window; if the average absolute error of the forecasting results meets a preset trigger condition, then online incremental learning and parameter optimization are performed on the target load forecasting model based on online elastic weight consolidation of multiple anchor points; wherein, the online elastic weight consolidation of multiple anchor points includes: establishing a set of historical state anchor points based on the model state of the target load forecasting model; when incremental learning is triggered, training is performed based on a loss function of new data batches and multiple anchor point constraints; the loss function of multiple anchor point constraints consists of the new data loss and the regularization loss of the set of historical state anchor points; and the model parameters of the target load forecasting model are updated online based on the moving exponential average method.

[0015] Optionally, the method further includes: visualizing the load forecast results.

[0016] Secondly, embodiments of the present invention also provide a cross-scene load prediction system based on portrait attention fusion and continuous learning, used to implement the cross-scene load prediction method based on portrait attention fusion and continuous learning provided in the embodiments of the present invention; the system includes: an acquisition module, a construction module, a matching module, a fusion module, a prediction module, and an optimization module; wherein, the acquisition module is used to acquire the running feedback data of the target scene and perform preprocessing and feature extraction to generate a running state feature vector, and calculate the credibility factor of the running state feature vector; the construction module is used to construct static attribute sub-vectors and behavior pattern sub-vectors based on the inherent physical characteristics and behavior patterns of the target scene, and construct a target scene portrait based on the static attribute sub-vectors and the behavior pattern sub-vectors; the matching module is used to calculate the matching degree score between the target scene portrait and each source scene portrait in the preset model library based on the credibility factor and a hybrid similarity matching algorithm of static similarity and morphological similarity. The system selects or integrates target load prediction models based on the credibility factor and the matching score. The preset model library includes different source scenario profiles and corresponding trained load prediction models. The target load prediction model is one load prediction model or an integrated model of multiple load prediction models in the preset model library. The fusion module is used to use the target scenario profile as a key-value pair and the running state feature vector as a query, and performs dynamic fusion based on an attention mechanism to obtain a fused feature vector. The attention mechanism is used to calculate the matching degree between the query and the key, and perform a weighted summation of the corresponding values ​​based on the matching degree. The prediction module is used to input the fused feature vector into the target load prediction model to predict the hot and cold loads of the target scenario and obtain the load prediction result. The optimization module is used to perform online incremental learning and parameter optimization of the target load prediction model based on the load prediction result and multi-anchor online elastic weight consolidation.

[0017] This invention provides a cross-scene load prediction method and system based on image attention fusion and continuous learning. By constructing a scene image that integrates static and behavioral pattern features and combining it with a hybrid similarity matching algorithm, it achieves accurate initialization for cross-scene model transfer. For the first time, it introduces architectural scene images as key-value pairs into the attention mechanism, dynamically fusing them with real-time running data, thus improving the collaborative perception and prediction capabilities of long-term characteristics and short-term states. Through a closed-loop optimization mechanism that integrates progressive fine-tuning, online incremental learning, and multi-anchor online elastic weight consolidation techniques, it ensures the continuous adaptability and stability of the model during long-term operation. This alleviates the technical problems of inaccurate model matching, significant deviations in models under dynamic operating conditions, and lack of adaptive optimization capabilities found in existing technologies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a cross-scene load prediction method based on profile attention fusion and continuous learning, provided in an embodiment of the present invention; Figure 2 A flowchart for constructing and matching a target scene profile is provided in an embodiment of the present invention; Figure 3 A flowchart of model selection and progressive hierarchical fine-tuning provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of multi-source information fusion and weight allocation provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a closed-loop feedback learning and online optimization mechanism provided in an embodiment of the present invention; Figure 6 A schematic diagram of a cross-scene load prediction system for portrait attention fusion and continuous learning provided for an embodiment of the invention; Figure 7 This is a schematic diagram of the overall architecture of another cross-scene load prediction system based on portrait attention fusion and continuous learning provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1

[0022] Figure 1 This is a flowchart of a cross-scene load prediction method based on portrait attention fusion and continuous learning according to an embodiment of the present invention. Figure 1 As shown, the method specifically includes the following steps: Step S102: Obtain the running feedback data of the target scene, perform preprocessing and feature extraction, generate the running state feature vector, and calculate the credibility factor of the running state feature vector.

[0023] Step S104: Construct static attribute sub-vectors and behavior pattern sub-vectors based on the inherent physical characteristics and behavior patterns of the target scene, and construct a target scene profile based on the static attribute sub-vectors and behavior pattern sub-vectors.

[0024] Step S106: Based on the credibility factor and a hybrid similarity matching algorithm of static similarity and morphological similarity, calculate the matching score between the target scene image and each source scene image in the preset model library, and select or integrate the target load prediction model based on the credibility factor and the matching score; the preset model library includes different source scene images and corresponding trained load prediction models, and the target load prediction model is a load prediction model in the preset model library or an integrated model of multiple load prediction models.

[0025] Step S108: The target scene image is used as a key-value pair, and the running state feature vector is used as a query. Dynamic fusion based on attention mechanism is performed to obtain a fused feature vector. The attention mechanism is used to calculate the matching degree between the query and the key, and the corresponding values ​​are weighted and summed based on attention weights according to the matching degree.

[0026] Step S110: Input the fused feature vector into the target load prediction model to predict the heating and cooling loads of the target scenario and obtain the load prediction results.

[0027] Step S112: Based on the load forecasting results and online elastic weight consolidation of multiple anchor points, perform online incremental learning and parameter optimization on the target load forecasting model.

[0028] Specifically, step S102 further includes the following steps: Step S1021: Obtain the operation feedback data of the target scenario; the operation feedback data includes environmental parameters, system status, control settings and real-time load.

[0029] Specifically, environmental parameters include outdoor temperature and solar radiation; system status includes the operating status and load rate of cold and heat source equipment, water pump frequency, and valve opening; control setpoints include indoor temperature setpoint and actual value, and operating mode indicator; and real-time load includes cooling load and / or heating load.

[0030] Step S1022: Perform data cleaning and multi-source time series alignment on the running feedback data to obtain preprocessed data.

[0031] Step S1023: Extract features from the preprocessed data to generate an operating status feature vector; the operating status feature vector includes: environmental features, system status features, control features, real-time load features, and time-series derived features.

[0032] Specifically, the data preprocessing process provided in this embodiment of the invention includes: 1. Data cleaning and repair: Short-term missing data handling: For data with ≤2 consecutive missing time points, linear interpolation is used to fill the missing data. Among them, x t This represents the runtime feedback data at time t.

[0033] Handling long-term missing data: For data with more than 2 consecutive missing time points, imputation is performed using the mean of historical data from the same period (e.g., the same time in the past N days). .

[0034] Outlier detection and handling: A sliding window Z-score method is used. Let the current window be... Calculate the window mean ( ) and standard deviation ( ).like If it is an anomaly, it is marked as an anomaly and the window mean is used. replace.

[0035] 2. Multi-source time series alignment: Unify all heterogeneous data streams to the basic time granularity of model prediction (e.g., 1 hour): High-frequency data (e.g., per minute): aggregated using arithmetic mean.

[0036] Low-frequency data (e.g., every 15 minutes): use linear interpolation to increase the frequency.

[0037] Finally, all data points are aligned based on the hourly timestamp to eliminate time shift errors.

[0038] 3. Operational status characteristic engineering: Based on the preprocessed data, a feature vector of the operating state with clear physical meaning is constructed. Its dimension is k. Core features may include (but are not limited to): environmental characteristics, comprehensive equipment operation index, control deviation, real-time load and its rate of change, moving average, trend characteristics, and periodic coding, etc. The following is an example: (1) Environmental characteristics include: outdoor dry bulb temperature outdoor relative humidity Solar radiation intensity .

[0039] (2) System state characteristics include: Equipment Comprehensive Operation Index Calculations are based on the operating status and load rate of key equipment (chillers, boilers, and pumps). Among them, The device status is 0 / 1. The load factor is 0-1. Weights are based on the device's rated power; Average valve opening Arithmetic mean of the opening degree of the hot / cold water valves; Average pump frequency .

[0040] (3) Regulatory characteristics include: Indoor temperature deviation: In the formula, To preset the indoor temperature, The actual measured temperature indoors; Working mode indicator : Weekdays are 1, weekends are 0; Special time markers : 1 for holidays or event days, 0 otherwise.

[0041] (4) Real-time load characteristics include: Current cooling load: Current heat load: ; Load change rate Current load L t Compared with the previous time step load L t-1 The difference .

[0042] (5) Time-series derived features include: Moving average characteristics: Moving average of load over the past 3 hours: ; Trend characteristics: The slope of the load over the past 6 hours was calculated based on linear regression. ;in, Indicates average time. Indicates average load; Periodic encoding: Sine and cosine encoding is used to avoid the influence of numerical magnitude. , , hour represents the number of hours.

[0043] Step S1024: Calculate the confidence factor based on the number of valid data points and the proportion of outliers in the running status feature vector; the formula for calculating the confidence factor includes: ; In the formula, α is the credibility factor.

[0044] Specifically, the confidence factor α ranges from (0,1) and is used to dynamically evaluate the input quality, thereby triggering adjustments to the backend fusion and prediction strategies.

[0045] In an optional embodiment of the present invention, the method further includes: adaptively adjusting the dimension of the running state feature vector based on a credibility factor; specifically including: If the credibility factor is not less than the first preset threshold, then the complete dimension of the running state feature vector is retained; If the credibility factor is less than the first preset threshold and greater than or equal to the second preset threshold, then the dimensionality reduction processing is performed on the running state feature vector; If the credibility factor is less than the second preset threshold, a data quality alarm signal will be issued.

[0046] For example, when When the value is ≥0.8, it is considered to have high confidence, and the complete running state feature vector is used for fusion and prediction. When 0.5≤ When the value is less than 0.8, it is considered to have medium confidence. The system automatically reduces the dimensionality of the operating status feature vector to a subset of core features (such as environmental parameters, basic load, and status of major equipment) to improve the robustness of the system. when When the value is less than 0.5, a data quality alarm is issued, which may trigger the model to revert to a more conservative prediction mode, and prompts the operation and maintenance personnel to intervene.

[0047] Figure 2 This is a flowchart of a target scene profile construction and matching process provided by an embodiment of the present invention. Figure 2 As shown, specifically, step S104 further includes the following steps: Step S1041: Based on the basic information of the target scenario, the thermal performance of the building envelope, the configuration parameters of the cooling and heating system, and the control and metering methods, construct a static attribute sub-vector; the thermal performance of the building envelope includes the building thermal inertia index.

[0048] Step S1042: Based on load and energy consumption time series patterns, personnel activity and space usage patterns, equipment system operation strategies and environmental response patterns, construct behavioral pattern sub-vectors; the environmental response patterns include energy consumption response patterns differentiated based on load-temperature sensitivity coefficients.

[0049] Specifically, this invention selects four types of static attributes that have the most significant impact on building heating and cooling loads, including basic building information, thermal performance of the building envelope, configuration parameters of the heating and cooling systems, and control and metering methods. All indicators are standardized to form an m-dimensional real-value vector. This invention particularly selects building physical characteristic indicators that have a significant impact on building heating and cooling loads and are rarely involved in other general load forecasting fields. For example, the building thermal inertia index reflects the building's ability to store and release heat, and has a significant impact on the phase delay of the load. Incorporating it into the static attribute vector can effectively distinguish buildings that appear similar but have vastly different thermal response characteristics, which is one of the keys to achieving accurate matching.

[0050] In one optional implementation, static attribute subvectors It can cover the following major categories of features (including but not limited to): (1) Basic building information: such as building type (code), year, building area, volume, shape coefficient, etc.

[0051] (2) Thermal performance of building envelope: such as comprehensive thermal inertia index, design heat consumption index, window-to-wall ratio for each orientation, heat transfer coefficient of exterior wall / exterior window / roof, airtightness index, etc. Among them, the building thermal inertia index is a static characteristic that is not widely used in load forecasting and migration in this field. It reflects the building's ability to store and release heat. The introduction of this index in this invention can more accurately characterize the delayed response characteristics of buildings to temperature fluctuations, thereby better distinguishing buildings with similar structures but different thermal responses in scene matching.

[0052] (3) Configuration parameters of cooling and heating systems: such as heat source network type, heat exchanger configuration parameters, circulating water pump parameters, indoor terminal system type and parameters, etc.

[0053] (4) Control and metering methods: such as system control method, climate compensation function, indoor temperature control method, heat metering method, etc.

[0054] This invention explores behavioral patterns from four dimensions: energy consumption, personnel, equipment, and environmental response. Unlike simply analyzing load curve morphology, this invention delves into the intrinsic laws governing building energy consumption behavior. For example, the energy consumption response pattern differentiated by the load-temperature sensitivity coefficient quantifies the load change caused by a unit temperature change, directly reflecting the combined effect of building envelope performance and internal heat gain characteristics. Introducing this coefficient greatly enhances the physical meaning and recognizability of the behavioral pattern profile. The extracted features are then formed into an n-dimensional vector.

[0055] In one optional implementation, behavior pattern subvectors This can be obtained by feature extraction from historical operational data, including (but not limited to): (1) Load and energy consumption time series pattern: Extract the typical daily load curve features of weekdays / weekends (e.g., by approximating PAA dimension reduction through segmented aggregation), weekly cycle intensity, daily load curve shape features (e.g., dynamic time warping (DTW) distance from the standard curve), load factor and fluctuation variance, etc.

[0056] (2) Personnel activity and space usage patterns: For example, generate personnel density curves based on sensor data and extract features (such as maximum / nighttime on-duty rate).

[0057] (3) Equipment system operation strategy: For example, extract time-series features such as the start and end dates of the heating season, daily start and stop times, and water temperature adjustment curves.

[0058] (4) Environmental response mode: For example, calculate the load-temperature sensitivity coefficient, the holiday load attenuation coefficient, etc. The load-temperature sensitivity coefficient is a dynamic feature that is not widely used in this field. It quantifies the sensitivity of building load to changes in outdoor temperature. The introduction of this coefficient in this invention can effectively distinguish the energy consumption response modes of different buildings under the same climatic conditions, and significantly improve the identification and matching accuracy of behavioral pattern profiles.

[0059] Specifically, such as Figure 2 As shown, the matching score calculation process in step S106 includes the following steps: Step S1061: Calculate the static similarity between the target scene image and the corresponding static attribute sub-vectors of each source scene image in the preset model library.

[0060] In obtaining target scene profile Then, its comparison with the source scene image is calculated. The degree of matching. Among them, Represents static attribute subvectors, This represents a behavior pattern subvector, where the subscript 'target' indicates the target scene and the subscript 'source' indicates the source scene.

[0061] Optionally, the static similarity S can be calculated using cosine similarity. S : .

[0062] Step S1062: Based on the dynamic time warping distance, calculate the morphological similarity between the corresponding behavioral pattern sub-vectors of the target scene image and each source scene image.

[0063] Specifically, the dynamic time warping (DTW) distance D is calculated from the dimensionality-reduced typical behavioral sequences (such as daily load curves) and converted into morphological similarity. .

[0064] Step S1063: Determine the matching weight based on the credibility factor.

[0065] For example, matching weight ,in, and The preset base weight and minimum guarantee weight (e.g.) =0.5, =0.7). The design intent is to appropriately increase the weight of morphological similarity when the historical operational data of the target scene is of high quality; and to rely more on the similarity of static physical attributes when the data is unreliable. This adaptive mechanism significantly improves the robustness of the matching.

[0066] Step S1064: Based on the matching weight, the static similarity and morphological similarity are weighted and fused to obtain the matching score between the target scene image and each source scene image in the preset model library.

[0067] For example, the matching score S total The calculation formula is: .

[0068] Figure 3 This is a flowchart illustrating model selection and progressive hierarchical fine-tuning according to an embodiment of the present invention. For example... Figure 3 As shown, specifically, the target load forecasting model selection process in step S106 includes: Step S1065: Based on the matching score and the preset matching threshold, select multiple candidate models from the preset model library.

[0069] For example, if the preset matching threshold is 0.7, then the matching score S will be used for filtering. total Models with a value ≥0.7 are considered candidate models.

[0070] Step S1066: Based on the matching score, historical performance score, and model robustness score of each candidate model, calculate the multi-dimensional score for each candidate model. Specifically, the multi-dimensional score calculation formula includes: ; in, It is the multi-dimensional score of the i-th candidate model. It is the matching score of the i-th candidate model. It is the historical performance score (MAPE normalized) of the i-th candidate model in the source scene. It is the robustness score of the i-th candidate model, based on the amount of training data and the stability of the architecture.

[0071] Step S1067: If the confidence factor is not less than the third preset threshold, then the candidate model with the highest multi-dimensional score is determined as the target load prediction model.

[0072] For example, the third preset threshold is 0.8, which indicates high confidence ( ): Select the single model with the highest multi-dimensional score as the target load prediction model.

[0073] Step S1068: If the credibility factor is less than the third preset threshold and greater than or equal to the fourth preset threshold, then the multi-dimensional scores will be ranked in the top N. top The candidate models were determined as base models, and the corresponding ensemble weights were calculated based on the multi-dimensional scores of each base model. Multiple base models were then integrated based on these ensemble weights to obtain the target load prediction model; N top It is an integer greater than 1.

[0074] For example, the fourth preset threshold is 0.5, N top A value of 3 indicates medium confidence ( The Top-3 ensemble model is used, with the weights of each model allocated according to its softmax score. The top three models are selected. Calculate the ensemble weights: Final prediction output: This is to reduce the risk of choosing a single model.

[0075] Preferably, such as Figure 3 As shown, the method provided in this embodiment of the invention further includes: progressively fine-tuning the target load prediction model in a hierarchical manner. Specifically, this embodiment of the invention uses recent data (e.g., 14-30 days) of the target scenario for fine-tuning. To balance "rapid adaptation" and "avoidance of forgetting," this invention employs a three-stage progressive fine-tuning approach: 1. Phase One (Rapid Adaptation): Only fine-tune the output layer to quickly align with the dimensions of the target scene; 2. Phase Two (Feature Calibration): Unfreeze parts of the high-level network, introduce EWC regularization, and the loss function is: .in, These are the parameters of the pre-trained model. Fisher information for parameters (indicating parameter importance). This represents the regularization strength. This stage allows the model to adjust high-level feature representations, but constrains them from deviating from the general knowledge already learned from the source scene.

[0076] 3. Phase Three (Global Fine-tuning): Unfreeze the entire network and increase... The value is used for more refined but more constrained optimization.

[0077] Figure 4 This is a schematic diagram of multi-source information fusion and weight allocation provided by an embodiment of the present invention. Figure 4 As shown, step S108 further includes the following steps: Step S1081: Map the target scene profile, the historical operation feature sequence of the target scene, and the operation status feature vector to a unified dimension through a projection matrix to form a feature matrix.

[0078] Specifically, the static attribute subvector behavioral pattern subvectors Historical operational characteristic sequence and the current running state feature vector The feature matrix H is formed by mapping the projection matrix to a unified dimension. , where b is the bias term.

[0079] Step S1082: Generate a query vector based on the running state feature vector, generate a key vector and a value vector based on the feature matrix, calculate the matching degree between the query vector and each key vector, and calculate the corresponding attention weight.

[0080] Specifically, using the one that best represents the instantaneous state of the system. Generate query vector Q ( ), generate key vector K from all projected features ( ) and value vector V ( Calculate the matching degree between Q and each K, and obtain the attention weights through Softmax. This means that, for different real-time states, the system will automatically allocate different levels of attention to different long-term characteristics (such as building insulation performance and human activity patterns).

[0081] Then calculate the attention weights. This indicates the degree of attention paid to features from different sources in the current state: ; In the formula, d k The feature dimensions represent the query vector and key vector.

[0082] Step S1083: The value vector is weighted and summed based on the attention weights to obtain the fused value vector.

[0083] Specifically, the fused value vector: .

[0084] Step S1084: Concatenate the fused value vector with the running state feature vector to obtain the fused feature vector. Specifically, the fused feature vector... .

[0085] In this invention, the dynamic fusion step is the core of improving the model's dynamic perception capability. Its innovative architecture breaks away from the traditional approach of simply splicing portrait features with real-time features, instead creating a portrait of the target scene... This is considered a "knowledge base" describing the long-term, inherent characteristics of a building, and is used as the key and value in the attention mechanism. When the input is a feature vector representing the current instantaneous state of the building... When the query is used as a Query, the attention mechanism calculates the relevance between the Query and each Key (i.e., a feature dimension of the profile) in the knowledge base, and aggregates the Value accordingly with weights. This means that under different real-time operating conditions (e.g., high-density daytime periods vs. low-power nighttime periods), the model will automatically adjust its attention to different long-term characteristics such as building insulation performance and human activity patterns, thereby achieving true context-aware prediction.

[0086] Figure 5 This is a schematic diagram of a closed-loop feedback learning and online optimization mechanism provided by an embodiment of the present invention. Figure 5 As shown, step S112 further includes the following steps: Step S1121: Based on the average absolute error of the load prediction results monitored within the window using a sliding window.

[0087] Specifically, a sliding window error monitoring method is used, with the sliding window size defined as W (e.g., W = 168 hours, i.e., one week's worth of hourly data), and the mean absolute error (MAPE) of prediction within the window is calculated: ; in, This is the actual load. It is a predicted load.

[0088] Step S1122: If the mean absolute error of the prediction meets the preset triggering condition, then based on the online elastic weight consolidation of multiple anchor points, online incremental learning and parameter optimization are performed on the target load prediction model.

[0089] Preferably, the preset triggering conditions include: incremental learning can be triggered when one of the following conditions is met: (1) Continuous high triggering: The average absolute error of prediction exceeds the threshold (e.g., 15%) for N consecutive time periods. (2) Trend drift trigger: The average error of the prediction mean absolute error of multiple sliding windows in the recent period has increased by more than a certain percentage (e.g., 20%) compared with the historical benchmark. (3) Event-driven trigger: The event ends when a special event (such as equipment failure or major control) is detected.

[0090] Preferably, multi-anchor point online elastic weight consolidation includes: Based on the model state of the target load prediction model, establish a set of historical state anchor points; When incremental learning is triggered, training is performed based on a loss function that is based on new data batches and multiple anchor constraints; the loss function with multiple anchor constraints consists of the new data loss and the regularization loss of the historical state anchor set. The model parameters of the target load prediction model are updated online based on the moving exponential average method.

[0091] The online incremental learning and parameter optimization method provided in this invention aims to solve the performance degradation problem caused by slow or sudden changes in the scene during long-term operation, and to address the "catastrophic forgetting" and "stage-based knowledge loss" problems commonly found in existing online learning methods. The core of this method lies in optimizing online weight consolidation. Traditional Elastic Weight Consolidation (EWC) and its online version (Online-EWC) typically use only a single initially trained model parameter set. This serves as an anchor point to prevent the forgetting of initial knowledge. However, during long-term, continuous online learning, the model will gradually learn and form multiple important knowledge states at different stages in the target scenario (e.g., adapting to summer operating patterns, adapting to new weekday scheduling strategies, etc.). By only fixing the initial anchor point, the model may overwrite or forget these equally important stage-level knowledge learned during the evolution of the target scenario when adapting to subsequent new changes.

[0092] To address this problem, this invention proposes and implements a multi-anchor point online elastic weight consolidation technique: (1) Definition and maintenance of anchor point set: First, dynamically maintain a set of historically important state anchor points: ; Each anchor point is a triple, where: It is the first A snapshot of model parameters saved at a historically significant moment. Yes, it corresponds to The parameter importance matrix (usually a diagonal Fisher information matrix) is used to measure the importance of each parameter to the model's prediction in this state. This is the creation timestamp of the anchor point.

[0093] Anchor point creation timing: It is not created periodically, but after the model performance has significantly and stably improved, or after it is confirmed that the model has successfully learned and adapted to a new and stable operating mode (for example, by determining through a trigger mechanism that a learning cycle has been completed and the performance has met the target), the current model state is created as a new anchor point.

[0094] (2) Loss function for multi-anchor constraints: When incremental learning is triggered, based on the new data batch Perform training. Total loss function. It consists of the new data loss and the regularization loss based on all historical anchors: ; in, The loss is based on the new batch of data (e.g., mean squared error), K is the number of anchor points in the current anchor point set, and N is the total number of model parameters. It is the i-th parameter of the model that needs to be updated. It is the parameter value corresponding to the k-th anchor point. It is the Fisher information (importance) of the i-th parameter in the k-th anchor point, which is a scalar. It is the regularization strength coefficient for the k-th anchor point, which can be adjusted based on the "age" or importance of the anchor point (e.g., closer anchor points may be assigned a slightly larger value). ).

[0095] The core function of this loss function is to strive for a good fit to new data. At the same time, it forces the new parameters of the model. We must not deviate too much from any important historical state. Each anchor point is determined by its Fisher information content. The deviation behavior is subject to "flexible" constraints, with stronger constraints on more important parameters. In this way, the model's knowledge updates are confined to a "knowledge retention space" jointly defined by all key historical experiences, achieving collaborative protection of multidimensional historical knowledge.

[0096] (3) Online estimation and updating of Fisher information: Furthermore, this invention continuously estimates the importance (Fisher information) of current model parameters online for potential future creation of new anchor points. It employs a moving exponential average method for updates. ; in, It is the latest estimate of the importance of the i-th parameter in the current model. This is the forgetting factor (usually a small value such as 0.001). This is the loss calculated based on the most recent data. This step continues in the background to ensure that the importance estimate of the anchor point remains correlated with the current data distribution.

[0097] Specifically, in the method provided by this embodiment of the invention, to ensure the continuity of the real-time prediction service, the training process is carried out in an independent "shadow model" copy: Asynchronous training: When the triggering condition is met, the parameters of the current master prediction model are copied to the "shadow model", and the above-mentioned multi-anchor Online-EWC algorithm is used on the shadow model to train based on the new data and anchor set.

[0098] Smooth parameter synchronization: After the shadow model is trained, its parameters are synchronized. With the main model parameters Synchronize after smooth blending, rather than directly replacing: ; in, The mixing coefficient (e.g., 0.1) is used to avoid drastic fluctuations in prediction results due to parameter abrupt changes.

[0099] Management of anchor point sets and buffers: Set the maximum number of anchor points (e.g., 5). When the number of anchor points exceeds At that time, based on the anchor point "age" ( ) and average importance Calculate the elimination score and remove the anchor point with the lowest score: ; in, It is the elimination score at the kth anchor point. It is the average importance of the k-th anchor point. It is the current timestamp. It is the creation timestamp of the kth anchor point. It is the smoothing constant (taking the smallest positive number).

[0100] The size of the circular data buffer is dynamically adjusted based on the available system memory, and a minimum update interval is set. (For example, This prevents excessive and frequent learning from consuming too many resources.

[0101] Specifically, the method provided in this embodiment of the invention further includes: visualizing the load forecast results. Specifically, this embodiment of the invention also provides a graphical interface that displays forecast curves and confidence intervals, performance reports (such as MAPE), energy consumption analysis reports, and supports user intervention and configuration.

[0102] The method provided by this invention achieves accurate initialization for cross-scene model transfer by constructing a scene profile that integrates static and behavioral pattern features and combining it with a hybrid similarity matching algorithm. It is the first to introduce the building scene profile as a key-value pair into an attention mechanism, dynamically fusing it with real-time operational data to improve the collaborative perception and prediction capabilities of long-term characteristics and short-term states. Through a closed-loop optimization mechanism that integrates progressive fine-tuning, online incremental learning, and multi-anchor point online elastic weight consolidation techniques, it ensures the model's continuous adaptability and stability during long-term operation. This system and method provide an efficient, robust, and scalable solution for intelligent load forecasting of integrated energy systems in commercial buildings.

[0103] Example 2

[0104] Figure 6 This is a schematic diagram of a cross-scene load prediction system based on portrait attention fusion and continuous learning, according to an embodiment of the invention. Figure 6As shown, the system includes: an acquisition module 10, a construction module 20, a matching module 30, a fusion module 40, a prediction module 50, and an optimization module 60.

[0105] Specifically, the acquisition module 10 is used to acquire the running feedback data of the target scene, perform preprocessing and feature extraction, generate a running state feature vector, and calculate the credibility factor of the running state feature vector. Module 20 is used to construct static attribute sub-vectors and behavior pattern sub-vectors based on the inherent physical characteristics and behavior patterns of the target scene, and to construct a target scene profile based on the static attribute sub-vectors and behavior pattern sub-vectors. The matching module 30 is used to calculate the matching score between the target scene image and each source scene image in the preset model library based on the confidence factor and the hybrid similarity matching algorithm of static similarity and morphological similarity, and to select or integrate the target load prediction model based on the confidence factor and the matching score. The preset model library includes different source scene images and corresponding trained load prediction models. The target load prediction model is a load prediction model in the preset model library or an integrated model of multiple load prediction models. The fusion module 40 is used to dynamically fuse the target scene profile as a key-value pair and the running state feature vector as a query based on an attention mechanism to obtain a fused feature vector. The attention mechanism is used to calculate the matching degree between the query and the key, and to perform a weighted summation of the corresponding values ​​based on the attention weights according to the matching degree. The prediction module 50 is used to input the fused feature vector into the target load prediction model to predict the heating and cooling loads of the target scenario and obtain the load prediction results. The optimization module 60 is used to perform online incremental learning and parameter optimization of the target load forecasting model based on load forecasting results and online elastic weight consolidation of multiple anchor points.

[0106] Specifically, the acquisition module 10 is also used to: acquire operational feedback data of the target scenario; the operational feedback data includes environmental parameters, system status, control setpoints, and real-time load; perform data cleaning and multi-source time series alignment on the operational feedback data to obtain preprocessed data; extract features from the preprocessed data to generate an operational status feature vector; the operational status feature vector includes: environmental features, system status features, control features, real-time load features, and time-series derived features; and calculate a reliability factor based on the number of valid data points and the proportion of outliers in the operational status feature vector.

[0107] Specifically, the acquisition module 10 is also used to: adaptively adjust the dimension of the running state feature vector based on the credibility factor.

[0108] Specifically, module 20 is also used to: construct a static attribute sub-vector based on the basic information of the target scenario, the thermal performance of the building envelope, the configuration parameters of the cooling and heating system, and the control and metering methods; the thermal performance of the building envelope includes the building thermal inertia index; construct a behavioral pattern sub-vector based on the load and energy consumption time series pattern, the human activity and space usage pattern, the equipment system operation strategy, and the environmental response pattern; the environmental response pattern includes the energy consumption response pattern distinguished by the load-temperature sensitivity coefficient.

[0109] Specifically, the matching module 30 is also used to: calculate the static similarity between the target scene image and the corresponding static attribute sub-vectors of each source scene image in the preset model library; calculate the morphological similarity between the target scene image and the corresponding behavioral pattern sub-vectors of each source scene image based on the dynamic time warping distance; determine the matching weight based on the credibility factor; and perform weighted fusion of static similarity and morphological similarity based on the matching weight to obtain the matching score between the target scene image and each source scene image in the preset model library.

[0110] Specifically, the matching module 30 is also used to: filter multiple candidate models in a preset model library based on the matching score and a preset matching threshold; calculate a multi-dimensional score for each candidate model based on its matching score, historical performance score, and model robustness score; if the confidence factor is not less than a third preset threshold, then the candidate model with the highest multi-dimensional score is determined as the target load prediction model; if the confidence factor is less than the third preset threshold but greater than or equal to a fourth preset threshold, then the candidate model with the highest multi-dimensional score in the top N... top The candidate models were determined as base models, and the corresponding ensemble weights were calculated based on the multi-dimensional scores of each base model. Multiple base models were then integrated based on these ensemble weights to obtain the target load prediction model; N top It is an integer greater than 1.

[0111] Specifically, the fusion module 40 is also used to: map the target scene profile, the historical operation feature sequence of the target scene, and the operation state feature vector to a unified dimension through a projection matrix to form a feature matrix; generate a query vector based on the operation state feature vector, generate a key vector and a value vector based on the feature matrix, calculate the matching degree between the query vector and each key vector, and calculate the corresponding attention weight; perform a weighted summation of the value vector based on the attention weight to obtain a fused value vector; and concatenate the fused value vector with the operation state feature vector to obtain a fused feature vector.

[0112] Specifically, the optimization module 60 is also used to: monitor the average absolute error of the load forecast result within the window based on the sliding window; if the average absolute error of the forecast meets the preset triggering conditions, then perform online incremental learning and parameter optimization of the target load forecast model based on online elastic weight consolidation of multiple anchor points; wherein, online elastic weight consolidation of multiple anchor points includes: establishing a set of historical state anchor points based on the model state of the target load forecast model; when incremental learning is triggered, training is performed based on the loss function of new data batches and multiple anchor point constraints; the loss function of multiple anchor points consists of the new data loss and the regularization loss of the set of historical state anchor points; and the model parameters of the target load forecast model are updated online based on the moving exponential average method.

[0113] Example 3

[0114] Figure 7 This is a schematic diagram of the overall architecture of another cross-scene load prediction system based on portrait attention fusion and continuous learning, provided by an embodiment of the present invention. Figure 7 As shown, the system includes the following modules: 1. Multi-scenario pre-trained model library management module: This is used to store and maintain pre-trained load prediction models for various typical commercial building scenarios. The preferred model is a neural network structure that can capture local correlations and long-term dependencies of sequences, such as a CNN-LSTM hybrid network.

[0115] 2. Scene profile building and matching module: Used to build scene profiles specifically describing the building's heating and cooling load characteristics for the target scene and source scenes in the model library, and to calculate their matching degree. Scene Profile Static attribute subvectors characterizing the inherent physical properties of a building and behavioral pattern subvectors representing the regularity of building use Composition, that is .

[0116] Among them, static attribute sub-vectors The data constructed based on Building Information Modeling (BIM) and design drawings must include indicators reflecting the thermal performance of the building envelope, such as the window-to-wall ratio for each orientation, the heat transfer coefficient of the exterior walls / windows, and the building thermal inertia index reflecting the building's thermal response characteristics; behavioral pattern sub-vectors It is obtained by time-series pattern mining of historical operating data, and must include indicators that quantify the building load response characteristics to the external environment, such as the load-temperature sensitivity coefficient.

[0117] The scene profile construction and matching module uses a hybrid similarity algorithm to calculate the directional similarity between the target scene and the source scene on the static attribute vectors. (e.g., cosine similarity), and morphological similarity on typical behavioral sequences after dimensionality reduction (e.g., typical daily load curves). (e.g., similarity based on Dynamic Time Warping (DTW) distance transformation), and through adaptive weights The total matching score is obtained by fusion. Weight With real-time data credibility factor Dynamic association.

[0118] 3. Run the feedback data collection and extraction module: Used to collect, process, and extract key operational feature vectors from the target building's energy system in real time. The data covers environmental parameters, system status, control setpoints, and real-time load. One of the core innovations of this module is the introduction of a dynamic data quality assessment mechanism, which calculates a reliability factor α to trigger different fusion and prediction strategies subsequently.

[0119] 4. Multi-source information fusion and weight allocation module: This module is designed to achieve collaborative perception of long-term building scene characteristics and short-term operational status. Its core lies in employing an attention mechanism and innovatively incorporating static attribute sub-vectors of the scene profile. and behavior pattern subvectors As the key and value in the attention mechanism, the feature vector will be processed in real time. As a query, by calculating the matching degree between the query and the key, attention weights are dynamically assigned to different scene profile features, and finally a context-aware feature vector that integrates long-term knowledge (scene profile) and short-term state (running data) is generated as the input of the prediction model.

[0120] 5. Model selection and adaptive fine-tuning module: Based on total matching score Based on data reliability α, a single optimal model is selected from the pre-trained model library, or multiple models are integrated. When fine-tuning using initial data from the target scenario, a progressive hierarchical strategy is adopted, and Elastic Weight Consolidation (EWC) regularization is introduced to prevent catastrophic forgetting of pre-trained knowledge.

[0121] 6. Closed-loop feedback learning and online optimization module: After model deployment, predictive performance is continuously monitored and incremental learning is triggered. Its key innovation lies in employing a multi-anchor online elastic weight consolidation optimization technique. Unlike traditional elastic weight consolidation techniques that only fix initial parameters, this technique maintains a set containing multiple historically important parameter states (anchors). Each anchor records the parameter value and its importance (Fisher information) at a significant evolutionary stage. During incremental learning updates, the loss function includes a regularization term for all historical anchors, constraining new parameters from deviating excessively from each historically important state. This effectively prevents the model from forgetting previously learned key knowledge when adapting to new data, achieving continuous, smooth, and robust model evolution.

[0122] 7. Visualized Decision Support Module: It is used to provide users with a graphical interface to display prediction results, performance evaluation, energy consumption analysis and optimization suggestions.

[0123] As described above, the embodiments of the present invention provide a cross-scene load prediction method and system for portrait attention fusion and continuous learning, which has the following technical effects compared with the prior art: (1) High matching accuracy and excellent migration starting point: By introducing static physical properties (such as thermal inertia index) and behavioral indicators (such as load-temperature sensitivity) unique to the building field, a scene profile is constructed, and a two-dimensional hybrid similarity matching is adopted to achieve cross-scene model matching that is closer to the building load formation mechanism, which fundamentally improves the initial migration accuracy.

[0124] (2) Strong dynamic perception capability and more realistic prediction: By introducing scene profile as a key-value into the attention mechanism, the prediction model can dynamically pay attention to different long-term characteristics of buildings according to the real-time operating status, which significantly improves the model's perception and prediction accuracy under dynamic control conditions.

[0125] (3) Long-term evolution is robust and knowledge is not forgotten: Through the online elastic weight consolidation and optimization mechanism with multiple anchor points, the model can effectively retain important knowledge at each stage during continuous learning, which solves the problem of "catastrophic forgetting" that is prone to occur in traditional online learning and ensures the stability and sustainability of the model's long-term performance.

[0126] (4) The system has strong adaptability and good engineering applicability: The system is based on data credibility. The adaptive decision chain can dynamically adjust matching weights, model selection, and fusion strategies based on data quality, thereby improving the system's robustness and practicality in complex engineering environments.

[0127] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0128] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A cross-scene load prediction method based on portrait attention fusion and continuous learning, characterized in that, The method includes: The system acquires operational feedback data of the target scenario, performs preprocessing and feature extraction, generates operational status feature vectors, and calculates the credibility factor of the operational status feature vectors. Static attribute sub-vectors and behavioral pattern sub-vectors are constructed based on the inherent physical characteristics and behavioral patterns of the target scene, and a target scene profile is constructed based on the static attribute sub-vectors and the behavioral pattern sub-vectors. Based on the credibility factor and a hybrid similarity matching algorithm combining static similarity and morphological similarity, the matching score between the target scene image and each source scene image in the preset model library is calculated. Based on the credibility factor and the matching score, a target load prediction model is selected or integrated. The preset model library includes different source scene images and corresponding trained load prediction models. The target load prediction model is one load prediction model or an integrated model of multiple load prediction models in the preset model library. The target scene profile is used as a key-value pair, and the running state feature vector is used as a query. Dynamic fusion based on an attention mechanism is performed to obtain a fused feature vector. The attention mechanism is used to calculate the matching degree between the query and the key, and the corresponding values ​​are weighted and summed based on the attention weights according to the matching degree. The fused feature vector is input into the target load prediction model to predict the heating and cooling loads of the target scenario, and the load prediction result is obtained. Based on the load forecasting results and online elastic weight consolidation at multiple anchor points, the target load forecasting model is subjected to online incremental learning and parameter optimization. The process involves acquiring and preprocessing operational feedback data from the target scenario, extracting features to generate operational state feature vectors, and calculating the reliability factor of these vectors. Obtain operational feedback data for the target scenario; the operational feedback data includes environmental parameters, system status, control settings, and real-time load. The operational feedback data is cleaned and aligned using multiple source time series to obtain preprocessed data. Feature extraction is performed on the preprocessed data to generate an operating state feature vector; the operating state feature vector includes: environmental features, system state features, control features, real-time load features, and time-series derived features; The credibility factor is calculated based on the number of valid data points and the proportion of outliers in the operating status feature vector. Static attribute sub-vectors and behavioral pattern sub-vectors are constructed based on the inherent physical characteristics and behavioral patterns of the target scene, including: Based on the basic information of the target scenario, the thermal performance of the building envelope, the configuration parameters of the cooling and heating system, and the control and metering methods, a static attribute sub-vector is constructed; the thermal performance of the building envelope includes the building thermal inertia index. Based on load and energy consumption time-series patterns, personnel activity and space usage patterns, equipment system operation strategies, and environmental response patterns, behavioral pattern sub-vectors are constructed; the environmental response patterns include energy consumption response patterns differentiated based on load-temperature sensitivity coefficients.

2. The method according to claim 1, characterized in that: The method further includes: adaptively adjusting the dimension of the running state feature vector based on the credibility factor; including: If the credibility factor is not less than the first preset threshold, then the complete dimension of the running state feature vector is retained; If the credibility factor is less than the first preset threshold and greater than or equal to the second preset threshold, then the running state feature vector is subjected to dimensionality reduction processing. If the credibility factor is less than the second preset threshold, a data quality alarm signal is issued.

3. The method according to claim 1, characterized in that: Based on the aforementioned credibility factor and a hybrid similarity matching algorithm combining static similarity and morphological similarity, the matching score between the target scene image and each source scene image in the preset model library is calculated, including: Calculate the static similarity between the target scene profile and the corresponding static attribute sub-vectors of each source scene profile in the preset model library; Based on the dynamic time warping distance, the morphological similarity between the target scene profile and the corresponding behavioral pattern sub-vectors of each source scene profile is calculated. Based on the credibility factor, the matching weight is determined; Based on the matching weight, the static similarity and the morphological similarity are weighted and fused to obtain the matching score between the target scene image and each source scene image in the preset model library.

4. The method according to claim 1, characterized in that: Based on the credibility factor and the matching score, a target load prediction model is selected or integrated, including: Based on the matching score and the preset matching threshold, multiple candidate models are selected from the preset model library; A multi-dimensional score is calculated for each candidate model based on its matching score, historical performance score, and model robustness score. If the confidence factor is not less than the third preset threshold, then the candidate model with the highest multi-dimensional score is determined as the target load prediction model; If the credibility factor is less than the third preset threshold and greater than or equal to the fourth preset threshold, then the multi-dimensional score will be ranked in the top N. top The candidate models are determined as base models, and the corresponding ensemble weights are calculated based on the multi-dimensional scores of each base model. Multiple base models are then integrated based on these ensemble weights to obtain the target load prediction model; N top It is an integer greater than 1.

5. The method according to claim 1, characterized in that: Using the target scene profile as a key-value pair and the running state feature vector as a query, dynamic fusion based on an attention mechanism is performed to obtain a fused feature vector, including: The target scene profile, the historical operation feature sequence of the target scene, and the operation state feature vector are mapped to a unified dimension through a projection matrix to form a feature matrix; A query vector is generated based on the running state feature vector, a key vector and a value vector are generated based on the feature matrix, the matching degree between the query vector and each key vector is calculated, and the corresponding attention weight is calculated. The value vector is weighted and summed based on the attention weights to obtain a fused value vector. The fused value vector is concatenated with the running state feature vector to obtain the fused feature vector.

6. The method according to claim 1, characterized in that: Based on the load forecasting results and multi-anchor point online elastic weight consolidation, the target load forecasting model undergoes online incremental learning and parameter optimization, including: The average absolute error of the load prediction results within the window is monitored based on a sliding window. If the predicted mean absolute error meets a preset triggering condition, then based on multi-anchor point online elastic weight consolidation, online incremental learning and parameter optimization are performed on the target load prediction model; wherein, the multi-anchor point online elastic weight consolidation includes: Based on the model state of the target load prediction model, a set of historical state anchor points is established; When incremental learning is triggered, training is performed based on a loss function that is based on new data batches and multiple anchor constraints; the loss function that is based on multiple anchor constraints consists of new data loss and regularization loss of the set of historical state anchors. The model parameters of the target load prediction model are updated online based on the moving exponential average method.

7. The method according to claim 1, characterized in that: The method further includes: visualizing the load forecast results.

8. A cross-scene load prediction system based on portrait attention fusion and continuous learning, characterized in that, The system is used to implement the cross-scene load prediction method for portrait attention fusion and continuous learning as described in any one of claims 1-7; the system includes: an acquisition module, a construction module, a matching module, a fusion module, a prediction module, and an optimization module; wherein... The acquisition module is used to acquire the running feedback data of the target scene, perform preprocessing and feature extraction, generate a running state feature vector, and calculate the credibility factor of the running state feature vector. The construction module is used to construct static attribute sub-vectors and behavior pattern sub-vectors based on the inherent physical characteristics and behavior patterns of the target scene, and to construct a target scene profile based on the static attribute sub-vectors and the behavior pattern sub-vectors. The matching module is used to calculate the matching score between the target scene image and each source scene image in the preset model library based on the credibility factor and a hybrid similarity matching algorithm of static similarity and morphological similarity, and to select or integrate a target load prediction model based on the credibility factor and the matching score; the preset model library includes different source scene images and corresponding trained load prediction models, and the target load prediction model is one load prediction model or an integrated model of multiple load prediction models in the preset model library; The fusion module is used to take the target scene profile as a key-value pair and the running state feature vector as a query, and perform dynamic fusion based on an attention mechanism to obtain a fused feature vector; the attention mechanism is used to calculate the matching degree between the query and the key, and perform a weighted summation of the corresponding values ​​based on attention weights according to the matching degree. The prediction module is used to input the fused feature vector into the target load prediction model to predict the heating and cooling loads of the target scenario and obtain the load prediction result. The optimization module is used to perform online incremental learning and parameter optimization of the target load prediction model based on the load prediction results and online elastic weight consolidation of multiple anchor points.

Images