Energy management method and system combining big data with behavior prediction

Abstract

The application discloses an energy management method and system combining big data behavior prediction, relates to the technical field of intelligent energy management systems, and comprises the following steps: setting a preset detection time period; detecting the regulation and control behaviors of multiple regulation and control sides (users, systems and management), outputting corresponding data packets when at least two regulation and control sides are triggered; performing feature vectorization processing on the data packets into feature vectors, importing a prediction model for prediction; and connecting an energy management system to perform energy management according to the regulation and control prediction results. The application solves the technical problem that in the energy prediction and management scene, the traditional data processing mode is difficult to accurately perceive the regulation and control intention and reconcile the behavior conflicts, leading to low matching degree of energy scheduling and actual demand, and failing to meet the efficient energy management demand, and achieves the technical effects of accurately perceiving the regulation and control intention, reconciling the behavior conflicts, improving the matching degree of energy scheduling and demand, and realizing efficient energy management.

Description

Technical Field

[0001] This invention relates to the field of intelligent energy management system technology, and in particular to an energy management method and system that combines big data to predict regulatory behavior. Background Technology

[0002] In energy management, accurate prediction of regulatory behavior is crucial for improving energy efficiency and optimizing management decisions. Current energy management systems often rely on data from a single regulatory body or simple data fusion, lacking collaborative analysis of user, system, and management behaviors. Traditional methods fail to effectively integrate data and accurately predict regulatory intentions when faced with complex interactions across multiple stakeholders, resulting in low matching between energy dispatch and actual demand, hindering scientific and efficient energy management. Due to insufficient collaboration of multi-stakeholder behavioral data and the lack of hierarchical prediction mechanisms, traditional approaches are ill-suited to complex energy regulation scenarios and cannot meet the demands for accurate prediction and management. Summary of the Invention

[0003] This application provides an energy management method and system that combines big data to predict regulatory behavior, in order to solve the technical problem that traditional data processing methods are unable to accurately perceive regulatory intentions and reconcile behavioral conflicts in energy prediction and management scenarios, resulting in a low degree of matching between energy dispatch and actual demand, and failing to meet the needs of efficient energy management.

[0004] The first aspect of this application provides an energy management method combining big data-driven regulation behavior prediction. The method includes: setting a preset detection time period; detecting regulation behavior on multiple regulation sides within the preset detection time period; if at least two of the multiple regulation sides are detected to trigger regulation behavior, outputting at least two sets of regulation behavior data packets corresponding to the at least two regulation sides, wherein the multiple regulation sides include user regulation sides, system regulation sides, and management regulation sides; performing feature vectorization processing on the at least two sets of regulation behavior data packets, outputting at least two sets of feature vectors; importing the at least two sets of feature vectors into a fusion behavior prediction model for prediction, outputting regulation prediction results; and connecting to an energy management system to perform energy management based on the regulation prediction results.

[0005] A second aspect of this application provides an energy management system that combines big data-driven regulation behavior prediction. The system includes: a time period setting module for setting a preset detection time period; a behavior detection execution module for detecting regulation behavior on multiple regulation sides within the preset detection time period; if at least two of the multiple regulation sides are detected to trigger regulation behavior, outputting at least two sets of regulation behavior data packets corresponding to the at least two regulation sides, wherein the multiple regulation sides include user regulation sides, system regulation sides, and management regulation sides; a vectorization processing module for performing feature vectorization processing on the at least two sets of regulation behavior data packets, outputting at least two sets of feature vectors, importing the at least two sets of feature vectors into a fusion behavior prediction model for prediction, and outputting regulation prediction results; and an energy management execution module for connecting to the energy management system and executing energy management based on the regulation prediction results.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] This application collects control behavior data from the user side, system side, and management side in energy management scenarios. After feature extraction, standardization, and embedding of side-specific information, feature vectors are generated. These vectors are then used in conjunction with an intent-aware, three-level fusion module to determine the control behavior based on the user's intent. Figure 1 Consistency indicators predict regulatory behavior under different collaborative scenarios. By combining the amplification and demodulation strategies of each module to adjust the prediction results, the system can accurately predict multi-sided regulatory behavior, providing a reliable basis for the energy management system to execute strategies, improving the scientific nature and efficiency of energy management, and achieving the technical effect of accurately perceiving regulatory intentions, reconciling behavioral conflicts, improving the matching degree between energy dispatch and demand, and realizing efficient energy management. Attached Figure Description

[0008] 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.

[0009] Figure 1 This is a flowchart illustrating the energy management method that combines big data-driven behavioral prediction provided in the embodiments of this application.

[0010] Figure 2 This is a schematic diagram of the structure of an energy management system that combines big data-driven behavior prediction, as provided in an embodiment of this application.

[0011] Figure labeling: Time period setting module 1, Behavior detection execution module 2, Vectorization processing module 3, Energy management execution module 4. Detailed Implementation

[0012] This application provides an energy management method and system that combines big data to predict regulatory behavior, in order to solve the technical problem that traditional data processing methods are unable to accurately perceive regulatory intentions and reconcile behavioral conflicts in energy prediction and management scenarios, resulting in a low degree of matching between energy dispatch and actual demand, and failing to meet the needs of efficient energy management.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, an energy management method combining big data-driven behavioral prediction is described, wherein the method includes:

[0016] Step A100: Set the preset detection time period.

[0017] Specifically, the preset detection time period needs to be determined based on the actual needs of the energy management scenario. For example, it can be set to 15 minutes, 1 hour, or 1 day. The specific duration can be flexibly adjusted by those skilled in the art based on the response frequency and data update requirements of the energy system.

[0018] By setting a preset detection time period, this solution can achieve regular collection and phased processing of multi-control side behavior data. With the help of a standardized time window, it provides a unified time benchmark for subsequent feature vectorization, fusion prediction and energy management execution, effectively improving the efficiency and quality of data collection, reducing invalid computational overhead, and laying the foundation for accurately capturing multi-side behavior coordination patterns and improving the accuracy of energy control prediction.

[0019] Step A200: Detect control behavior of multiple control sides within the preset detection time period. If at least two of the multiple control sides are detected to trigger control behavior, output at least two sets of control behavior data packets corresponding to the at least two control sides. The multiple control sides include user control side, system control side and management control side.

[0020] In this embodiment, the control side refers to the three types of entities involved in energy management, including the user control side, the system control side, and the management control side. Control behavior refers to energy management-related behaviors triggered by the control side within a preset detection period, such as the user side adjusting the electricity consumption mode, the system side executing automatic control commands, and the management side issuing policy updates.

[0021] Optionally, at the start of the preset detection time period, the main scheduler sends behavior detection requests to multiple control sides to collect control behavior data packets from the user side, system side, and management side. It performs content change analysis on the data packets. If no control behavior is triggered, it skips the current time period and enters the next time period. If it is triggered by a single control side, it caches the corresponding control behavior data packet. The specific steps are explained in detail in A210-A230.

[0022] Furthermore, for example, in a certain industrial park management system, user-side electricity meters upload data every 15 minutes, system-side load data is updated every 5 minutes, and management policies are released daily. The inconsistent time bases of these three systems cause timestamp discrepancies of 3-5 minutes during multi-side behavioral correlation analysis, leading to increased error rates in prediction models due to time-series inconsistencies. To address this, a multi-side data acquisition process with a unified time window is constructed using a 1-hour preset time period and a synchronous API acquisition mechanism, resolving the coordination issues related to time series and data types.

[0023] At the start of the preset detection period (e.g., 1-hour cycle, triggered on the hour), the main scheduler, as the core control unit, synchronously initiates behavior detection requests to the three types of control sides through a standardized API interface:

[0024] User control side: Call the smart meter API to collect power consumption (e.g., current value 2.5kW, fluctuation ±1.2kW compared to the previous period) and equipment start / stop status (e.g., machine tool M01 in the workshop starts at 14:00). The data is returned in structured JSON format, including device ID, behavior type (power adjustment / equipment start / stop), and timestamp (accurate to the moment).

[0025] System control side: Access the power monitoring system API to obtain grid voltage (e.g., 10kV bus voltage fluctuation ±0.3kV) and load rate (e.g., a feeder load increases from 60% to 75%). The data is transmitted in real-time data stream format, with additional monitoring point number, behavior type (load warning / voltage anomaly), and timestamp.

[0026] Management and control side: Connect to the policy release platform API to capture energy efficiency indicator adjustments (e.g., the energy consumption threshold per unit in the industrial park is changed from 1.2 kWh / Reduced to 1.1kWh / Peak-valley electricity price updates (e.g., adjusting the off-peak electricity price from 0.3 yuan / kWh to 0.28 yuan / kWh), the data is returned in the form of instruction text, including instruction type, controlled object (all workshops / specific production line), and timestamp (the time when the policy takes effect).

[0027] After the three types of requests are initiated, the main scheduler listens for responses through an asynchronous parallel mechanism to ensure that all data packets are collected within 5 seconds (user-side response time ≤ 2 seconds, system-side ≤ 1.5 seconds, and management-side ≤ 1.5 seconds), forming a dataset for the time period containing timestamps, side identifiers, and raw data fields. After collection, the system first verifies time synchronization: if the timestamp deviation of the three sides is ≤ 5 seconds, it is determined to be valid synchronization within the time period; if the deviation exceeds 5 seconds, a time alignment algorithm is triggered, such as interpolating and correcting the system-side and management-side data based on the user-side timestamp, to ensure the temporal consistency of subsequent analysis.

[0028] When at least two control sides are detected to trigger actions during the time period, such as user-side power change ≥5% and system-side load rate change ≥3%, or management side releases new policies and user-side equipment starts and stops, it is immediately marked as a multi-side collaborative event. The original data packets of the corresponding sides (preserving API return format characteristics) and time-aligned metadata are output. This mechanism provides a multi-source data foundation with unified time series and type adaptation for subsequent feature extraction and model input.

[0029] By synchronously collecting and collaboratively triggering behavioral data from the user side, system side, and management side within a preset detection period, multi-dimensional perception of energy regulation behavior is achieved. Valid data packets are output only when there is a substantial correlation between the data from multiple sides. This avoids the one-sidedness of data from a single side and provides a high-quality collaborative dataset for subsequent feature vectorization and fusion prediction, thereby improving the scientific nature and response efficiency of energy management.

[0030] Step A300: Perform feature vectorization processing on the at least two sets of regulation behavior data packets, output at least two sets of feature vectors, import the at least two sets of feature vectors into the fusion behavior prediction model for prediction, and output the regulation prediction result.

[0031] In this embodiment of the application, the prediction model is a hierarchical model used to analyze the feature vectors of multiple control sides and output the control prediction results.

[0032] In one embodiment of this application, behavior fields are extracted from at least two sets of control behavior data packets to obtain behavior type, control object, control content and behavior timestamp, and then standardized; control side embedding vector combination is introduced to standardize behavior fields and output at least two sets of feature vectors. The specific steps are described in detail in A310-A330.

[0033] The fusion behavior prediction model comprises an intent-aware module and three fusion modules. Feature vectors are input into the intent-aware module for analysis to obtain regulatory intent. Figure 1 Consistency index; based on this index, activate the corresponding fusion module to predict and output the corresponding regulation prediction results. The specific steps are explained in detail in A340-A370.

[0034] Step A400: Connect to the energy management system and perform energy management based on the aforementioned regulation and forecast results.

[0035] In this embodiment of the application, the energy management system is a system for receiving the regulation prediction results output by the fusion behavior prediction model and performing energy management accordingly.

[0036] Specifically, firstly, the output of the fusion behavior prediction model is obtained: the semantic consistency of multi-sided feature vectors is analyzed through the intent perception module, and the corresponding fusion module is activated to generate the regulation prediction result. After receiving the regulation prediction result, the energy management system first performs strategy parsing and device mapping: the abstract prediction is transformed into specific execution instructions, such as adjusting the power threshold of the smart meter corresponding to a 1.5kW increase in power on the user side, and triggering the compensation device of the feeder corresponding to the activation of reactive power compensation on the system side.

[0037] Then, the priority-based execution phase begins: for highly consistent prediction results (such as regulatory intentions), the priority-based execution phase proceeds. Figure 1 With a consistency index ≥ 0.7, the system executes the weighted fusion enhancement strategy with a response speed of 0.5 seconds, such as amplifying the energy saving to 15%. For compromise strategies in conflict scenarios, for example, if a user increases the load by 5% and starts energy storage, the system executes the response step by step according to the response priority (user demand 0.6, system stability 0.4) to ensure that the load fluctuation is controlled within ±8%.

[0038] Finally, a closed-loop correction is achieved through real-time monitoring and feedback: the system continuously collects real-time data after execution, such as the user's actual power consumption and system load rate, and compares it with the prediction results to generate a deviation value (e.g., allowable error ±3%). If the deviation exceeds the threshold, the model is triggered to recalculate. For example, when it is detected that the load rate still reaches 85% after the energy-saving strategy is implemented, the system automatically calls the second fusion module to generate a supplementary strategy, such as adding a 5% load reduction, forming a dynamic optimization link of prediction-execution-correction.

[0039] By deeply coupling the fusion prediction results with the equipment control logic of the energy management system, a complete closed loop from multi-source data perception to precise strategy execution is constructed. This process solves the static execution defects of traditional systems through a three-level mechanism of model prediction results driving strategy generation, dynamic priority scheduling, and real-time feedback correction. This improves the matching degree between the actual effect of energy regulation strategy and prediction target, and enhances the system's adaptability and management efficiency in complex energy scenarios.

[0040] Furthermore, step A200 in the method provided in this application embodiment includes:

[0041] A210: At the start of the preset detection time period, the main scheduler sends behavior detection requests to the multiple control sides to collect user-side control behavior data packets, system control side control behavior data packets, and management control side control behavior data packets.

[0042] A220: Perform content change analysis on the user-side control behavior data packet, system control-side control behavior data packet, and management control-side control behavior data packet. If it is detected that no control behavior has been triggered in the multiple control sides, skip the current preset detection time period and enter the next preset detection time period.

[0043] A230: If it is detected that the multiple control sides trigger the control behavior by a single control side, the control behavior data packet corresponding to the single control side is cached.

[0044] Specifically, at the start of the preset detection period, the main scheduler, as the core control unit of the system, first sends behavior detection requests synchronously to the user side, system side, and management side. For example, taking a preset period of 1 hour as an example, at each hour, the main scheduler initiates data acquisition requests through the API interface to the user-side smart meters (collecting data such as power consumption and equipment start / stop), the system-side power monitoring system (collecting parameters such as grid voltage and load rate), and the management-side policy release platform (collecting instructions such as energy efficiency index adjustment and peak-valley electricity price update). Within 5 seconds, it completes the collection of three types of control behavior data packets, forming a raw dataset containing fields such as timestamp, behavior type, and parameter value.

[0045] After data collection, content change analysis is performed on the three types of data packets. Specifically, this is achieved by comparing the difference between the current period's data and the previous period's data: assuming the user-side power fluctuation is <5%, the system-side load rate change is <3%, and there are no new instructions pushed by the management side (the specific difference threshold can be determined by those skilled in the art based on the actual needs of the energy management scenario), it is determined that no control behavior has been triggered. In this case, the current period is skipped to avoid meaningless processing of static data. If only the user-side power fluctuation is ≥10%, such as when a user turns on a high-power appliance, but there is no change in the system-side and management-side data, it is determined that a single control side has triggered the data. The user-side data packet is then stored in a cache queue, the cache time is marked, and it awaits the combination of data from subsequent periods.

[0046] By using the timed request triggering of the main scheduler, the synchronous collection of multi-side data, and the intelligent filtering mechanism based on content changes, the resource consumption of invalid data processing is reduced. At the same time, by caching data from one side, the data foundation for cross-cycle multi-side behavior collaborative analysis is preserved. Thus, while improving operational efficiency, the foundation is laid for the accuracy of subsequent multi-source data fusion prediction.

[0047] Furthermore, step A240 in the method provided in this application embodiment includes:

[0048] A241: If no other control side triggering control behavior is detected at the end of the preset detection period, the energy management system performs energy management according to the control behavior data packet corresponding to the cache control side.

[0049] A242: When at the end of the preset detection time period, if at least one other control side different from the cache control side is detected to trigger a control behavior, extract the control behavior data packet corresponding to the cache control side and combine it with the control behavior data packet corresponding to the at least one other control side to output at least two sets of control behavior data packets.

[0050] In this embodiment of the application, the cached control side is the corresponding control side that caches the control behavior data packets corresponding to a single control side.

[0051] Optionally, after detecting the control behaviors of multiple control sides within the aforementioned preset detection time period, the corresponding triggering conditions are shown in Table 1. When only a single control side triggering behavior is detected within the preset detection time period (e.g., set to 1 hour), such as the management side issuing a 10% load reduction instruction for the afternoon session at 14:00, the system stores the management side data packet in the cache queue, marks the cache time as 14:00-15:00, and starts dynamic monitoring before the end of the cycle. At the end of the time period (15:00), the system first scans other control sides (user side, system side) for new triggering behaviors:

[0052] No other side triggering scenario: If the power consumption fluctuation on the user side is less than 5% and the load rate on the system side is stable, it is determined to be a valid instruction on one side. The energy management system directly executes the cached management side instruction, such as adjusting the power supply threshold of each area. This process does not require starting the fusion prediction module, and the computational resource consumption is reduced compared to the traditional full-process processing.

[0053] Other side-triggered scenarios: If a high-power device is detected to start in a workshop on the user side at 14:55, triggering a power change of ≥15%, the system immediately extracts the cached management side data packet and the newly collected user side data packet, combines them into a collaborative dataset of management side load reduction instruction + user side load increase requirement, and outputs it to subsequent feature vectorization processing.

[0054] By managing the cache of data packets on a single control side and triggering detection across cycles, the blindness of directly processing data from a single side in the traditional model is avoided. Furthermore, by delaying decision-making, data combination conditions are created for potential multi-side collaboration. This improves the efficiency of computing resource utilization while ensuring the immediacy of energy management strategies in single-side dominant scenarios and the accuracy in multi-side collaborative scenarios.

[0055] Table 1: Multi-side Data Acquisition and Triggering Conditions

[0056]

[0057] Furthermore, step A300 in the method provided in this application embodiment includes:

[0058] A310: Extract the behavior fields of the at least two sets of control behavior data packets respectively, and output the behavior type, control object, control content and behavior timestamp.

[0059] A320: Standardize the behavior type, control object, control content, and behavior timestamp to obtain at least two sets of standardized behavior fields.

[0060] A330: Introduce a control side embedding vector to combine the at least two sets of standardized behavioral fields and output at least two sets of feature vectors.

[0061] In this embodiment of the application, the control side embedding vector is a vector used to combine standardized behavioral fields. By introducing this vector, at least two sets of standardized behavioral fields can be combined to output at least two sets of feature vectors.

[0062] Specifically, firstly, the execution behavior field is extracted: core fields are parsed for the control behavior data packets from the user side, system side, and management side. Taking the user side as an example, the behavior type (e.g., device start / stop), control object (e.g., smart meter ID-001), control content (e.g., power increasing from 2.5kW to 3.8kW), and timestamp are extracted from the JSON format data packet; on the system side, the behavior type (e.g., voltage fluctuation), control object (e.g., feeder L-03), control content (e.g., voltage decreasing from 10.2kV to 9.8kV), and timestamp are extracted from the data stream; on the management side, the behavior type (e.g., energy efficiency index adjustment), control object (e.g., all users), control content (e.g., carbon emission standard decreasing from 0.5t / month to 0.4t / month), and timestamp (e.g., policy effective date) are extracted from the policy text.

[0063] Next, when performing feature vectorization processing on at least two sets of control behavior data packets, regular expression matching technology is first used to accurately extract fields such as behavior type, control object, control content, and timestamp from structured data (such as user-side JSON format data packets and system-side real-time data streams); for unstructured text data from the management side, semantic analysis is performed using natural language processing (NLP) technology to extract key information such as instruction type and control object, thereby achieving accurate extraction of fields from both structured and unstructured data.

[0064] Then, standardization is performed: the extracted fields are standardized in type and normalized in value. For example, behavior types are mapped to enumerated values, such as device start / stop → 01, voltage fluctuation → 02, energy efficiency adjustment → 03; control objects are encoded as unique identifiers, such as smart meter ID-001 → OBJ-001; control content values ​​are normalized according to the system range, such as power value divided by rated power, voltage value divided by standard voltage; timestamps are converted to relative time within a preset time period, such as converting to decimals of 0-1 within a 1-hour cycle. Taking user-side power data as an example, when the rated power is 5kW, 3.8kW is normalized to 0.76; under the standard value of 10kV system voltage, 9.8kV is normalized to 0.98. After standardization, the dimensions of multi-source data are unified into 4 core fields, ultimately yielding at least two sets of standardized behavior fields, improving data format consistency and laying the foundation for model input.

[0065] Finally, we introduce the side-specific embedding vector: Define one-hot encoded vectors for the user side, system side, and management side, such as user side → [1,0,0], system side → [0,1,0], and management side → [0,0,1]. Concatenate these vectors with the standardized behavioral fields to generate a multi-dimensional feature vector containing side-specific features. For example, a user-side data point, after processing, forms a feature vector: [01,OBJ-001,0.76,0.25,1,0,0], where the first four dimensions are the standardized behavioral fields, and the last three dimensions are the side-specific embedding vector. This vector not only preserves behavioral details but also explicitly expresses the data source, enabling the fusion model to distinguish the influence weights of different control sides.

[0066] Through the above steps, heterogeneous regulation behavior data from the user side, system side, and management side are transformed into feature vectors with unified structure and clear semantics. This solves the problems of chaotic data fields and implicit side information in traditional methods, providing high-quality feature representation for subsequent fusion behavior prediction models and significantly enhancing the accuracy and reliability of energy regulation behavior prediction.

[0067] Furthermore, step A300 in the method provided in this application embodiment includes:

[0068] A340: The fusion behavior prediction model includes an intent perception module, a first fusion behavior prediction module, a second fusion behavior prediction module, and a third fusion behavior prediction module.

[0069] A350: Input the at least two sets of feature vectors into the intent perception module for analysis to obtain the control intent. Figure 1 Consistency indicators, if the aforementioned regulatory intention Figure 1 If the consistency index is greater than or equal to the preset consistency index threshold, the first fusion behavior prediction module is activated to predict the at least two sets of feature vectors and output the first regulation prediction result.

[0070] A360: If the aforementioned regulation intention Figure 1 If the consistency index is less than the preset consistency index threshold and the return value is not empty, the second fusion behavior prediction module is activated to predict the at least two sets of feature vectors and outputs the second regulation prediction result.

[0071] A370: If the aforementioned regulation intention Figure 1 If the consistency index returns empty, the third fusion behavior prediction module is activated to predict the at least two sets of feature vectors and outputs the third regulation prediction result.

[0072] In this embodiment, the intent perception module is used to perform semantic analysis on each feature vector in at least two sets of feature vectors, obtain at least two sets of semantic vectors, calculate the semantic similarity of each of the at least two sets of semantic vectors, output the mean semantic similarity, and obtain the control intent based on the mean semantic similarity. Figure 1 The consistency index module. The fusion behavior prediction module is a model that includes an intent perception module, a first fusion behavior prediction module, a second fusion behavior prediction module, and a third fusion behavior prediction module.

[0073] Specifically, the core process of the fusion behavior prediction model is the semantic analysis of the intent perception module: at least two sets of feature vectors are input into this module, and natural language processing (NLP) technology is used to perform semantic parsing on the behavioral fields (such as power regulation, load warning) in each vector, generating corresponding semantic vectors, such as user-side vector [energy saving, equipment, +15%] and system-side vector [stability, power grid, -10%]. Then, the cosine similarity of each set of semantic vectors is calculated, and the mean semantic similarity is output as the control intent. Figure 1 Consistency index (value range 0-1). For example, when both the user side and the management side involve energy-saving semantics, the average similarity can reach above 0.8; if the user side is increasing load while the system side is reducing load, the similarity may be below 0.3.

[0074] Based on Figure 1 Consistency indicators, the model dynamically triggers different fusion modules:

[0075] In high consistency scenarios (indicators ≥ threshold, such as 0.7): the first fusion behavior prediction module obtains the first fusion feature vector by fusing at least two sets of feature vectors by weighted average, introduces the first amplification coefficient to enhance the features, and then performs energy regulation behavior prediction and outputs the first regulation prediction result. The specific steps are explained in detail in A351-A353.

[0076] In conflict or low consistency scenarios (0.2 < index < 0.7): The second fusion behavior prediction module first divides at least two groups of feature vectors into consistent and conflict feature vector groups. After weighted average fusion, the conflict group is reconciled using a conflict demodulation network. Based on the fused and reconciled feature vectors, the second regulation prediction result is output. The specific steps are explained in detail in A361-A365.

[0077] In scenarios where data is missing or unrelated (indicator = empty): the third fusion behavior prediction module predicts at least two sets of feature vectors respectively, and uses the output of each side's regulatory behavior as the third regulatory prediction result. The specific steps are explained in detail in A371-A372.

[0078] By constructing a hierarchical prediction model that includes intent perception and a three-level fusion module, based on the control intent... Figure 1 The consistency index dynamically adapts to different collaborative scenarios, enabling accurate prediction of multi-sided regulatory behaviors and improving the scientific nature and system stability of energy management strategies.

[0079] Furthermore, step A350 in the method provided in this application embodiment includes:

[0080] A351: Wherein, the first fusion behavior prediction module includes a first amplification coefficient.

[0081] A352: Activate the first fusion behavior prediction module to perform weighted average fusion of the at least two sets of feature vectors to obtain the first fusion feature vector.

[0082] A353: Introduce the first amplification factor to enhance the features of the first fused feature vector, output the first enhanced fused feature vector, perform energy regulation behavior prediction on the first enhanced fused feature vector, and output the first regulation prediction result.

[0083] Specifically, firstly, after the first fusion behavior prediction module is activated, it performs a weighted average fusion of at least two sets of feature vectors. Based on preset side weights, such as 0.4 for the user side, 0.3 for the system side, and 0.3 for the management side, the vectors are weighted and summed according to dimensions such as behavior type and controlled object to generate the first fusion feature vector. For example, after weighted fusion of the user-side feature vector [energy saving, air conditioning, +10%, 0.5, 1, 0, 0] and the management-side vector [energy saving, all users, -15%, 0.5, 0, 0, 1], the fusion vector [energy saving, air conditioning / all users, -5%, 0.5, 0.4, 0, 0.6] is obtained. This vector comprehensively reflects the synergy between the two sides in achieving energy-saving goals.

[0084] Subsequently, a first amplification factor (e.g., 1.2) is introduced to enhance the features of the fused vector. This is achieved by linearly amplifying the collaborative feature dimensions; for example, increasing the semantic weights related to energy conservation from -5% to -6%, thus highlighting multi-faceted meanings. Figure 1 The influence of consistency is considered, generating the first enhanced fusion feature vector. After enhancement, the proportion of collaborative features in the vector increases, improving the model's sensitivity to highly consistent scenarios. Finally, energy regulation behavior prediction is performed based on the enhanced vector, outputting the first regulation prediction result.

[0085] By combining weighted average fusion and feature enhancement strategies, the first fusion behavior prediction module achieves in-depth mining and enhanced expression of multi-sided highly consistent regulatory intentions, thereby improving the execution and synergistic effect of energy regulation strategies while ensuring prediction accuracy.

[0086] Furthermore, step A360 in the method provided in this application embodiment includes:

[0087] A361: Wherein, the second fusion behavior prediction module includes a conflict demodulation network.

[0088] A362: Activate the second fusion behavior prediction module to divide the at least two sets of feature vectors and output a consistent feature vector group and a conflicting feature vector group.

[0089] A363: Perform a weighted average fusion of the at least two sets of feature vectors to obtain a fused feature vector.

[0090] A364: The conflict demodulation network is introduced to harmonize the conflict feature vector group to obtain harmonized feature vectors, wherein the harmonization objectives include response priority and system load size.

[0091] A365: Based on the fused feature vector and the harmonic feature vector, predict energy regulation behavior and output a second regulation prediction result.

[0092] In this embodiment, the conflict demodulation network is a component in the second fusion behavior prediction module used to reconcile conflict feature vector groups.

[0093] In one embodiment, after the second fusion behavior prediction module is activated, it first performs consistency analysis on at least two sets of feature vectors: by using a preset conflict threshold, such as semantic similarity < 0.5 to determine a conflict, the feature vectors are divided into consistent feature vector groups (such as fields involving equipment operating status on both sides) and conflicting feature vector groups (such as opposing fields of user-side load increase +15% and system-side load decrease -10%). Taking an industrial park scenario as an example, the user-side feature vector is [load increase, machine tool, +15%, 0.6, 1, 0, 0], and the system-side feature vector is [load decrease, feeder, -10%, 0.6, 0, 1, 0]. After division, the consistent feature vector group is [equipment, time 0.6], and the conflicting feature vector group is [load increase +15%, load decrease -10%].

[0094] Subsequently, all original input feature vectors before partitioning, which are processed by the second fusion behavior prediction module, are obtained. This constitutes the initial set of at least two feature vectors used for consistency analysis and subsequent weighted average fusion operations. A weighted average fusion is performed on these vectors to generate a basic fusion feature vector, such as a comprehensive power change of +5%. Simultaneously, a conflict demodulation network is activated for the conflict feature vector groups. This network uses response priority (e.g., production demand priority 0.7, system stability priority 0.3) and system load size (e.g., triggering a load reduction priority strategy when the current load rate is >80%) as harmonization objectives, and reconstructs the conflict fields using a linear combination algorithm. For example, in the above conflict scenario, the demodulated output harmonized feature vector [power change +3%, priority 0.7:0.3] retains some user-side demand while keeping the system load within a safe threshold.

[0095] Ultimately, based on the combination of fused feature vectors and harmonic feature vectors, such as the base power +5% plus the harmonic version +3%, a total adjustment of +8% is formed, and the second regulation prediction result is output, thereby improving the system's ability to balance the demand for conflict regulation.

[0096] By structurally partitioning, weighted fusion, and directional harmonizing the conflict demodulation network of multi-sided feature vectors, the second fusion behavior prediction module achieves refined processing of regulation conflicts, maximizing the satisfaction of multi-sided needs while ensuring system stability, thereby improving the scientificity and reliability of energy management strategies in complex scenarios.

[0097] Furthermore, step A370 in the method provided in this application embodiment includes:

[0098] A371: Activate the third fusion behavior prediction module to predict the at least two sets of feature vectors respectively, and output the first-side regulation behavior and the second-side regulation behavior.

[0099] A372: The output of the first-side regulation behavior and the second-side regulation behavior is used as the third regulation prediction result.

[0100] Optionally, when the regulatory intention Figure 1 If the consistency index returns empty, meaning the average semantic similarity of the feature vectors on each side is lower than a preset threshold (e.g., <0.2), the third fusion behavior prediction module is activated and performs independent prediction processes on at least two sets of feature vectors. Taking user-side and system-side unrelated data as an example:

[0101] User-side feature vector processing: Taking [Device start / stop, Air conditioner, +1, 0.3, 1, 0, 0] as an example, the behavior type of device start / stop, the air conditioner control object, the +1 control content (indicating the on status), and the 0.3 timestamp (at 30% of the time period) are first extracted from the behavior field. After standardization, a feature vector is generated by combining it with the user-side embedded vector [1, 0, 0]. The module analyzes this vector using a time-series prediction algorithm. Based on the time-series correlation between air conditioner on-time and power changes in historical data, for example, if similar devices previously had an average power increase of 1.2-1.8kW when on-time at 30% of the time period, an independent prediction of a 1.5kW increase in user-side power is output. This algorithm achieves prediction by identifying periodic and trend features in the time series.

[0102] System-side feature vector processing: Taking [voltage fluctuation, feeder A, -3%, 0.3, 0, 1, 0] as an example, the voltage fluctuation behavior type, feeder A control object, -3% control content (voltage drop amplitude), and 0.3 timestamp are extracted, and a feature vector is generated by combining it with the system-side embedded vector [0, 1, 0]. The module analyzes through a load model, which has built-in voltage thresholds and load adjustment rules. For example, reactive power compensation is triggered when the voltage drop is ≥2%. Based on the current load rate of the feeder, assuming that the probability of the load rate exceeding 85% when the voltage drop is 3% in historical data is 90%, an independent strategy for the system side to activate the reactive power compensation device is output. The model realizes the prediction and control of system stability by simulating the voltage-load dynamic relationship.

[0103] The above-mentioned side prediction results are directly output as the third regulation prediction results in parallel without the need for fusion processing.

[0104] By independently analyzing and predicting the feature vectors of multiple control sides in parallel, the third fusion behavior prediction module maintains the effectiveness of prediction and the independence of strategy in scenarios with sparse or unrelated data. This avoids errors caused by ineffective fusion and ensures that the energy management system can quickly respond to the control needs of one side under complex operating conditions, thereby improving the robustness and practicality of the overall prediction model.

[0105] Furthermore, step A350 in the method provided in this application embodiment includes:

[0106] A354: The intent-aware module is used to perform semantic analysis on each feature vector in the at least two sets of feature vectors to obtain at least two sets of semantic vectors.

[0107] A355: Calculate semantic similarity for each of the at least two sets of semantic vectors, output the mean semantic similarity, and obtain the control intention based on the mean semantic similarity. Figure 1 Consistency indicators.

[0108] In one embodiment, the intent-aware module first performs vector-wise semantic analysis on at least two sets of feature vectors: using Natural Language Processing (NLP) technology, it performs word segmentation, part-of-speech tagging, and entity recognition on text fields such as behavior type and control content in the feature vectors. For example, the power adjustment +15% in the user-side feature vector is parsed into the semantic vector [user demand, increased load, 15%], the system-side load warning -10% is parsed into [system status, reduced load, 10%], and the management-side peak-valley electricity price adjustment is parsed into [policy instruction, price incentive, none]. These semantic fields are mapped into vectors in a high-dimensional space using a word vector model (such as Word2Vec), forming at least two sets of semantic vectors.

[0109] Subsequently, cosine similarity is calculated for each group of semantic vectors: taking the user-side and system-side vectors as an example, the cosine of the angle between them in the semantic space is calculated. If the value is 0.3 (indicating semantic opposition between load increase and load decrease), it is judged as low consistency; if the similarity between the management-side and user-side vectors is 0.8 (such as the association between energy saving and price incentives), it is judged as high consistency. The average of all pairwise similarity values ​​is then calculated, and the mean semantic similarity is output, which is used as the regulatory intention... Figure 1 Consistency index (value range 0-1).

[0110] Through semantic analysis and similarity calculation processes in the intent-aware module, the system can accurately quantify the degree of collaboration among multiple regulatory intentions, providing a decision-making basis for the fusion behavior prediction model. This step addresses the shortcomings of traditional methods in semantic association analysis of multi-source heterogeneous data, elevating the evaluation of regulatory intentions from simple numerical matching to a deep understanding at the semantic level. This improves the accuracy of multi-side behavior collaborative prediction and lays the foundation for the dynamic triggering of hierarchical fusion strategies.

[0111] In summary, the energy management method combining big data-driven behavioral prediction provided in this application has the following technical effects:

[0112] This application detects control behaviors from the user control side, system control side, and management control side within a preset detection time period, collects multi-side control behavior data packets, obtains feature vectors through feature vectorization processing, and imports them into a fusion behavior prediction model to analyze control intentions. Figure 1 The system uses consistency indicators to activate different modules for prediction based on the indicator results. It combines the control and prediction results output by each module to perform energy management, thereby realizing energy management that combines big data control behavior prediction. This makes the energy control prediction results more comprehensive and accurate, achieving the technical effect of accurately perceiving control intentions, reconciling behavioral conflicts, improving the matching degree between energy dispatch and demand, and realizing efficient energy management.

[0113] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides an energy management system that combines big data-driven behavioral prediction, the system comprising:

[0114] Time period setting module 1 is used to set a preset detection time period.

[0115] The behavior detection execution module 2 is used to detect the control behavior of multiple control sides within the preset detection time period. If at least two of the multiple control sides are detected to trigger control behavior, at least two sets of control behavior data packets corresponding to the at least two control sides are output. The multiple control sides include user control side, system control side and management control side.

[0116] The vectorization processing module 3 is used to perform feature vectorization processing on the at least two sets of regulation behavior data packets, output at least two sets of feature vectors, import the at least two sets of feature vectors into the fusion behavior prediction model for prediction, and output the regulation prediction result.

[0117] Energy management execution module 4 is used to connect to the energy management system to perform energy management based on the control and prediction results.

[0118] Furthermore, the behavior detection execution module 2 is used to perform the following steps:

[0119] At the start of the preset detection period, the main scheduler sends behavior detection requests to the multiple control sides, collecting user-side control behavior data packets, system-side control behavior data packets, and management-side control behavior data packets. It then performs content change analysis on these data packets. If no control behavior is detected among the multiple control sides, the current preset detection period is skipped, and the next preset detection period begins. If a single control side triggers a control behavior, the control behavior data packet corresponding to that single control side is cached.

[0120] Furthermore, the behavior detection execution module 2 is used to perform the following steps:

[0121] If no other control side triggering control behavior is detected at the end of the preset detection period, the energy management system performs energy management according to the control behavior data packet corresponding to the cache control side; if at least one other control side triggering control behavior is detected at the end of the preset detection period, the control behavior data packet corresponding to the cache control side is extracted and combined with the control behavior data packet corresponding to the at least one other control side, and at least two sets of control behavior data packets are output.

[0122] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0123] The behavior fields of the at least two sets of control behavior data packets are extracted respectively, and the behavior type, control object, control content and behavior timestamp are output. The behavior type, control object, control content and behavior timestamp are standardized to obtain at least two sets of standardized behavior fields. The control side embedding vector is introduced to combine the at least two sets of standardized behavior fields to output at least two sets of feature vectors.

[0124] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0125] The fusion behavior prediction model includes an intent perception module, a first fusion behavior prediction module, a second fusion behavior prediction module, and a third fusion behavior prediction module. The at least two sets of feature vectors are input into the intent perception module for analysis to obtain the regulatory intent. Figure 1 Consistency indicators, if the aforementioned regulatory intention Figure 1 If the consistency index is greater than or equal to a preset consistency index threshold, the first fusion behavior prediction module is activated to predict the at least two sets of feature vectors and outputs a first regulation prediction result; if the regulation intention is... Figure 1 If the consistency index is less than the preset consistency index threshold and the return value is not empty, the second fusion behavior prediction module is activated to predict the at least two sets of feature vectors and outputs the second regulation prediction result; if the regulation intention is... Figure 1 If the consistency index returns empty, the third fusion behavior prediction module is activated to predict the at least two sets of feature vectors and outputs the third regulation prediction result.

[0126] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0127] The first fusion behavior prediction module includes a first amplification coefficient; the first fusion behavior prediction module is activated to perform weighted average fusion of the at least two sets of feature vectors to obtain a first fusion feature vector; the first amplification coefficient is introduced to enhance the features of the first fusion feature vector, and a first enhanced fusion feature vector is output; energy regulation behavior prediction is performed on the first enhanced fusion feature vector, and a first regulation prediction result is output.

[0128] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0129] The second fusion behavior prediction module includes a conflict demodulation network; it activates the second fusion behavior prediction module to divide the at least two sets of feature vectors, outputting a consistent feature vector set and a conflict feature vector set; it performs a weighted average fusion of the at least two sets of feature vectors to obtain a fused feature vector; it introduces the conflict demodulation network to harmonize the conflict feature vector set to obtain a harmonized feature vector, wherein the harmonization objective includes response priority and system load size; and it performs energy regulation behavior prediction based on the fused feature vector and the harmonized feature vector, outputting a second regulation prediction result.

[0130] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0131] The third fusion behavior prediction module is activated to predict the at least two sets of feature vectors respectively, and outputs the first-side regulation behavior and the second-side regulation behavior; the output of the first-side regulation behavior and the second-side regulation behavior is the third regulation prediction result.

[0132] Furthermore, the vectorization processing module 3 is used to perform the following steps:

[0133] The intent perception module is used to perform semantic analysis on each feature vector in the at least two sets of feature vectors to obtain at least two sets of semantic vectors; calculate the semantic similarity of each of the at least two sets of semantic vectors, output the mean semantic similarity, and obtain the control intent based on the mean semantic similarity. Figure 1 Consistency indicators.

[0134] The energy management system combining big data regulation behavior prediction provided in the embodiments of the present invention can execute the energy management method combining big data regulation behavior prediction provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0135] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0136] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. Energy management method incorporating big data for behavior prediction, characterized in that, The method comprises: setting a preset detection time period; detecting the regulation behavior of multiple regulation sides in the preset detection time period, and if at least two regulation sides in the multiple regulation sides trigger the regulation behavior, outputting at least two groups of regulation behavior data packets corresponding to the at least two regulation sides, wherein the multiple regulation sides include user regulation sides, system regulation sides and management regulation sides; performing feature vectorization processing on the at least two groups of regulation behavior data packets, outputting at least two groups of feature vectors, and inputting the at least two groups of feature vectors into a fusion behavior prediction model for prediction, and outputting a regulation prediction result; connecting an energy management system to perform energy management according to the regulation prediction result; wherein the method of inputting the at least two groups of feature vectors into the fusion behavior prediction model for prediction and outputting the regulation prediction result comprises: wherein the fusion behavior prediction model comprises an intention perception module, a first fusion behavior prediction module, a second fusion behavior prediction module and a third fusion behavior prediction module; inputting the at least two groups of feature vectors into the intention perception module for analysis to obtain a regulation intention consistency index, if the regulation intention consistency index is greater than or equal to a preset consistency index threshold, activating the first fusion behavior prediction module to predict the at least two groups of feature vectors and outputting a first regulation prediction result; if the regulation intention consistency index is less than the preset consistency index threshold and returns not empty, activating the second fusion behavior prediction module to predict the at least two groups of feature vectors and outputting a second regulation prediction result; if the regulation intention consistency index returns empty, activating the third fusion behavior prediction module to predict the at least two groups of feature vectors and outputting a third regulation prediction result; wherein the method of activating the second fusion behavior prediction module to predict the at least two groups of feature vectors and outputting the second regulation prediction result comprises: wherein the second fusion behavior prediction module comprises a conflict resolution network; activating the second fusion behavior prediction module to divide the at least two groups of feature vectors, outputting a consistent feature vector group and a conflict feature vector group; performing weighted average fusion on the at least two groups of feature vectors to obtain a fusion feature vector; introducing the conflict resolution network to reconcile the conflict feature vector group to obtain a reconciliation feature vector, wherein the reconciliation target includes response priority and system load size; performing energy regulation behavior prediction based on the fusion feature vector and the reconciliation feature vector to output the second regulation prediction result.

2. The method of claim 1, wherein, The method of detecting the regulation behavior of multiple regulation sides in the preset detection time period comprises: at the beginning of the preset detection time period, a main scheduler sends a behavior detection request to the multiple regulation sides to collect user side regulation behavior data packets, system regulation side regulation behavior data packets and management regulation side regulation behavior data packets; performing content change analysis on the user side regulation behavior data packets, system regulation side regulation behavior data packets and management regulation side regulation behavior data packets, if it is detected that the multiple regulation sides do not trigger the regulation behavior, skipping the current preset detection time period and entering the next preset detection time period; If the multiple regulation sides are detected as a single regulation side, triggering the regulation behavior, the single regulation side corresponding regulation behavior data packet is cached.

3. The method of claim 2, wherein, After the single regulation side corresponding regulation behavior data packet is cached, the method further comprises: If no other regulation side different from the cached regulation side is detected to trigger the regulation behavior at the end of the preset detection time period, the energy management system executes energy management according to the regulation behavior data packet corresponding to the cached regulation side; If at least one other regulation side different from the cached regulation side is detected to trigger the regulation behavior at the end of the preset detection time period, the regulation behavior data packet corresponding to the cached regulation side and the regulation behavior data packet corresponding to the at least one other regulation side are combined, and at least two groups of regulation behavior data packets are output.

4. The method of claim 1, wherein, The method for feature vectorization processing of the at least two groups of regulation behavior data packets comprises: The behavior field of the at least two groups of regulation behavior data packets is extracted respectively, and the behavior type, control object, control content and behavior timestamp are output. The behavior type, control object, control content and behavior timestamp are standardized to obtain at least two groups of standardized behavior fields; The at least two groups of standardized behavior fields are combined by introducing regulation side embedding vectors, and at least two groups of feature vectors are output.

5. The method of claim 1, wherein, The first fusion behavior prediction module is activated to predict the at least two groups of feature vectors, and a first regulation prediction result is output, and the method comprises: The first fusion behavior prediction module comprises a first amplification coefficient. The first fusion behavior prediction module is activated to perform weighted average fusion on the at least two groups of feature vectors, and a first fusion feature vector is obtained. The first amplification coefficient is introduced to perform feature enhancement on the first fusion feature vector, and a first enhanced fusion feature vector is output. The first enhanced fusion feature vector is used for energy regulation behavior prediction, and a first regulation prediction result is output. The third fusion behavior prediction module is activated to predict the at least two groups of feature vectors, and a third regulation prediction result is output, and the method comprises:

6. The method of claim 1, wherein, The third fusion behavior prediction module is activated to predict the at least two groups of feature vectors respectively, and a first side regulation behavior and a second side regulation behavior are output. The first side regulation behavior and the second side regulation behavior are output as the third regulation prediction result. The intention perception module is used for semantic analysis of each feature vector in the at least two groups of feature vectors, and at least two groups of semantic vectors are obtained.

7. The method of claim 1, wherein, The semantic similarity of the at least two groups of semantic vectors is calculated respectively, and the semantic similarity average value is output. The semantic similarity average value is used to obtain a regulation intention consistency index. The system is used for implementing the energy management method for combined big data regulation behavior prediction according to any one of claims 1-7, and the system comprises:

8. An energy management system incorporating big data to forecast behavior for managing, characterized by, A time period setting module is used for setting a preset detection time period. ​ The behavior detection execution module is configured to perform behavior detection on the plurality of regulation sides within the preset detection time period, and if it is detected that at least two regulation sides among the plurality of regulation sides trigger regulation behavior, output at least two groups of regulation behavior data packets corresponding to the at least two regulation sides, wherein the plurality of regulation sides include a user regulation side, a system regulation side, and a management regulation side. The vectorization processing module is configured to perform feature vectorization processing on the at least two groups of regulation behavior data packets, output at least two groups of feature vectors, and import the at least two groups of feature vectors into a fusion behavior prediction model for prediction, and output a regulation prediction result. The energy management execution module is configured to connect an energy management system to perform energy management according to the regulation prediction result.

Images