Panoramic monitoring method and system based on light storage straight flexible, computer device and medium

Abstract

This invention relates to the field of power monitoring data processing technology, and particularly to a panoramic monitoring method, system, computer equipment, and medium based on photovoltaic-storage-direct-drive-flexible (PV-SHU) systems. The method includes: configuring multi-dimensional heterogeneous data sources covering PV-SHU-direct-drive-flexible equipment, building structures, the environment, and the power grid to construct a distributed panoramic sensing data acquisition layer; sequentially performing data cleaning, scale normalization, timestamp synchronization, and spatial registration operations on the acquired monitoring data to generate a standardized data set aligned in time and space; constructing and executing a collaborative data fusion computing pipeline from the edge to the cloud; and publishing the fusion features and indicators output from the collaborative fusion stage through a standardized application programming interface. This approach solves the technical problems of insufficient accuracy, real-time performance, and matching degree between system collaborative scheduling and carbon efficiency optimization requirements in existing technologies when facing the large-scale application of PV-SHU-direct-drive-flexible systems.

Description

Technical Field

[0001] This invention relates to the field of power monitoring data processing technology, and in particular to a panoramic monitoring method, system, computer equipment, and medium based on photovoltaic-storage-direct-flexible integrated circuits. Background Technology

[0002] With the increasingly in-depth integration of photovoltaic, energy storage, DC power distribution, and flexible power consumption (hereinafter referred to as "PV-storage-DC-flexible") technologies in buildings, buildings are transforming from traditional single energy terminals into new integrated energy nodes that combine "source, storage, load, and flexibility." This transformation places higher demands on the operational status perception, multi-source data processing, and collaborative optimization capabilities of their internal energy systems, necessitating the construction of a monitoring system capable of providing panoramic visibility, deep data fusion, and agile response to system status changes. In the field of data processing and status assessment, existing technologies already offer solutions based on multi-source data fusion. For example, existing technology (application publication number CN120541681A) discloses a "Multi-source Data Fusion Status Assessment Method and System for Distribution Substations Based on the Internet of Things." This solution collects multi-source data such as electrical, temperature, and vibration data from distribution substations, utilizes an edge-cloud collaborative computing architecture for data preprocessing and feature extraction, and performs health status assessment and trend prediction in the cloud based on dynamic fuzzy rules and an LSTM (Long Short-Term Memory) model. This scheme represents a data processing method for condition assessment using multi-source data in a specific scenario (power distribution substation).

[0003] However, applying such methods directly to building-based "PV-Storage-DC-Flexible" systems still faces specific technical challenges at the electrical digital data processing level. Firstly, at the data level, "PV-Storage-DC-Flexible" systems involve multi-dimensional heterogeneous data such as PV power generation, energy storage charging and discharging status, DC bus voltage, flexible load adjustment potential, building sub-item energy consumption, and indoor and outdoor environmental parameters. The data types, sampling frequencies, and data quality (such as the intermittency of PV power and the discontinuity of load data) are more complex than the relatively stable electrical data of distribution substations. Existing data processing methods lack a systematic fusion framework for this specific multimodal and multi-timescale data, leading to fragmented data value and difficulty in accurately depicting the building's panoramic operational status as a comprehensive energy node. Secondly, at the data processing method level, existing fusion algorithms suffer from insufficient accuracy or large computational delays when dealing with the real-time alignment and deep fusion of second-level / minute-level equipment data with hourly meteorological data and daily carbon emission factors in "PV-Storage-DC-Flexible" systems. This makes it difficult to meet the stringent requirements of data timeliness and consistency for advanced applications such as internal collaborative scheduling and real-time precise carbon efficiency management. Finally, at the data service level, a "gap" exists between the processed data and advanced analytical models such as the collaborative optimization of building energy systems and the interactive response with the power grid. Data processing results fail to be effectively transformed into standardized, high-quality inputs that can directly drive optimization decisions, resulting in a disrupted "data-model-application" link. Therefore, existing building monitoring technologies face technical challenges in matching the accuracy and real-time performance of multi-source heterogeneous data fusion with the needs of system collaborative scheduling and carbon efficiency optimization when facing the large-scale application of photovoltaic-storage-DC-flexible systems. Summary of the Invention

[0004] To address the aforementioned shortcomings or deficiencies, this invention provides a panoramic monitoring method, system, computer equipment, and medium based on photovoltaic-storage-direct-flexible systems. This invention can solve the technical problems of insufficient accuracy, real-time performance, and matching degree between system collaborative scheduling and carbon efficiency optimization requirements when facing the large-scale application of photovoltaic-storage-direct-flexible systems.

[0005] This invention provides a panoramic monitoring method based on optical storage, direct and flexible imaging, comprising: Configure multi-dimensional heterogeneous data sources covering photovoltaic-storage-direct-drive-flexible equipment, building bodies, environment, and power grid side, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer.

[0006] The monitoring data collected by the distributed panoramic perception data acquisition layer is sequentially cleaned, scale normalized, timestamp synchronized and spatially registered to generate a standardized data set that is spatiotemporally aligned.

[0007] A collaborative data fusion computing pipeline from edge to cloud is constructed and executed. This pipeline includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge, a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud, and a collaborative fusion stage that performs coupled computation on system operating status and energy and carbon indicators.

[0008] The fusion characteristics and indicators output from the collaborative integration phase will be published through standardized application programming interfaces.

[0009] According to a second aspect, the present invention provides a panoramic monitoring system based on optical storage, direct and flexible imaging, comprising: The data acquisition layer construction module is used to configure multi-dimensional heterogeneous data sources covering photovoltaic storage direct current and flexible equipment, building body, environment and power grid side, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer.

[0010] The data set generation module is used to perform data cleaning, scale normalization, timestamp synchronization and spatial registration operations on the monitoring data collected by the distributed panoramic perception data acquisition layer in sequence, so as to generate a standardized data set that is aligned in time and space.

[0011] The computational pipeline construction module is used to build and execute a collaborative data fusion computational pipeline from the edge to the cloud. The pipeline includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge, a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud, and a collaborative fusion stage that performs coupled computation on system operating status and energy and carbon indicators.

[0012] The panoramic monitoring results generation module is used to publish the fusion features and indicators output from the collaborative fusion phase through a standardized application programming interface.

[0013] According to a third aspect, the present invention provides a computer device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to execute any of the panoramic monitoring methods based on optical storage and flexible positioning in the embodiments of the present invention.

[0014] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute any of the panoramic monitoring methods based on optical storage, directivity, and flexibility in the embodiments of the present invention.

[0015] The present invention provides a panoramic monitoring method based on photovoltaic-storage-direct-drive-flexible (PV-SHU) and flexible-drive (LD-H) systems. This method is achieved through four core steps: panoramic perception construction, data standardization generation, collaborative fusion calculation, and standardized result release. Specifically, it configures multi-dimensional heterogeneous data sources covering PV-SHU-LD-drive equipment, building structures, the environment, and the power grid, and deploys edge data acquisition agents to construct a distributed panoramic perception data acquisition layer. This provides a comprehensive, traceable monitoring data foundation covering the entire spectrum of "source-grid-load-storage-loop" for subsequent fusion calculations. The monitoring data acquired by this acquisition layer undergoes sequential data cleaning, scale normalization, timestamp synchronization, and spatial registration operations to generate a spatiotemporally aligned standardized data set. This eliminates differences in units, sampling frequencies, and spatial reference systems among the multi-source heterogeneous data, providing a unified and standardized input for high-quality data fusion. Finally, it constructs and executes collaborative data collection from the edge to the cloud. The fusion computing pipeline achieves shallow fusion by performing adaptive filtering on real-time data streams at the edge to ensure real-time processing. It then performs deep learning feature extraction on cross-modal data in the cloud to achieve deep fusion and uncover complex correlations. Finally, it performs coupled calculations on system operating status and energy and carbon indicators to achieve collaborative fusion and generate fusion indicators that can directly serve optimization decisions. This forms a progressive fusion capability from real-time processing to deep analysis and then to application-oriented approaches. The fusion features and indicators output from the collaborative fusion stage are published through standardized application programming interfaces (APIs), achieving loose coupling and standardized integration between monitoring results and upper-level scheduling, control, and optimization applications.

[0016] In this technical solution, the present invention addresses the problem of insufficient fusion accuracy caused by incomplete data source coverage and spatiotemporal misalignment, as described in the background technology. By constructing a distributed panoramic perception data acquisition layer and performing strict spatiotemporal alignment operations on multi-source data, the completeness and consistency of input data are ensured from the data source and preprocessing levels, laying a solid foundation for improving fusion accuracy. Furthermore, addressing the problem of balancing real-time performance and accuracy due to the lack of a fusion architecture adapted to distributed, highly volatile data, the present invention designs a collaborative computing pipeline consisting of shallow edge-side fusion, deep cloud-side fusion, and collaborative fusion. This leverages edge computing to ensure low latency of real-time data streams. The initial processing is delayed, and cloud computing power is used to support deep correlation mining and complex index coupling calculation of cross-modal data, thereby achieving a unity of real-time data processing and analytical depth at the architecture level. Addressing the problem of low utilization efficiency of fusion results and inability to effectively support collaborative scheduling and carbon efficiency optimization due to the disconnect between monitoring data and advanced applications, this invention directly couples the calculation of energy and carbon indicators such as operating status, carbon emissions, and flexible regulation value during the collaborative fusion stage. The final results are then published through a standardized application programming interface, enabling the output fusion features and indicators to be directly and efficiently invoked by optimization models such as photovoltaic-storage-DC-flexible collaborative scheduling, building energy and carbon management, and source-grid-load-storage cluster control. Therefore, the technical solution of this invention, through the systematic synergy of the above steps, solves the technical problems of insufficient accuracy, real-time performance, and matching degree of system collaborative scheduling and carbon efficiency optimization requirements in the face of large-scale application of photovoltaic-storage-DC-flexible systems in existing technologies. This improves the accuracy, responsiveness, and support capability for low-carbon operation decisions in buildings and new power systems. Attached Figure Description

[0017] Figure 1 This is a flowchart of a panoramic monitoring method based on optical storage, direct and flexible transmission according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the technical solution implementation architecture applied to a large commercial complex in another embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a panoramic monitoring system based on optical storage and flexible positioning according to an embodiment of the present invention; Figure 4 This is a block diagram of a computer device for implementing embodiments of the present invention. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0019] During the development of this invention, researchers conducted numerous experiments and data analyses, revealing the intrinsic connections between multi-dimensional data such as the operational status of the photovoltaic-storage-direct current-flexible system, building energy load, environmental parameters, and grid interaction information. These data are not isolated but collectively constitute a key information network reflecting the dynamic balance and carbon flow status of the building's integrated energy nodes ("source-grid-load-storage"). The depth, accuracy, and real-time nature of this fusion directly determine the effectiveness of subsequent collaborative scheduling and carbon efficiency management decisions. Based on this relationship, this invention innovatively proposes this technical solution. Utilizing a distributed processing architecture combining edge computing and cloud collaboration, it constructs a three-layer progressive data fusion computing pipeline ("shallow-deep-collaborative"), combined with spatiotemporal alignment preprocessing of multi-source heterogeneous data and multiple verification mechanisms for results. This achieves accurate, real-time perception of the panoramic operational status of the photovoltaic-storage-direct current-flexible system and standardized output of fusion indicators, embodying the core concept of "data-driven, fusion perception, and service-oriented decision-making."

[0020] Specifically, through comparative experiments, the invention team discovered that traditional single-dimensional or simply superimposed building monitoring methods suffer from technical defects such as fragmented data sources, limited fusion methods, and disconnect from advanced applications. Their data acquisition fails to cover the core equipment and multi-dimensional environment of photovoltaic-storage-directed and flexible energy storage systems; data processing lacks effective alignment for heterogeneity and time delay; and data application remains at the level of status display rather than decision support. These technical defects prevent a comprehensive depiction of the building's operational status as an energy node, and the output data is difficult to use directly and reliably for photovoltaic-storage coordinated scheduling and carbon efficiency optimization. The panoramic monitoring and fusion method proposed in this invention improves the completeness and accuracy of system operational status perception; by deploying edge data acquisition agents and constructing a distributed perception layer, it enables the nearby and efficient acquisition of data from dispersed devices; by performing strict spatiotemporal alignment operations to generate standardized data sets, it ensures the consistency and comparability of multi-source heterogeneous data before fusion; and by constructing and executing a three-layer fusion pipeline from edge to cloud collaboration, combined with standardized application programming interfaces to publish fusion results, it ensures the efficient and accurate conversion of monitoring data into scheduling optimization instructions.

[0021] In this invention, the "PV-Storage-DC-Flexible" system refers to a new type of building energy system that integrates photovoltaic power generation, energy storage, DC power distribution, and flexible power consumption technologies, achieving coordinated operation of these four elements. Its core components include: distributed photovoltaic power generation units, electrochemical or physical energy storage devices, a low-voltage DC power distribution network, and power-consuming terminals (i.e., flexible loads) with adjustable power capabilities. This system aims to achieve efficient absorption and storage of renewable energy within the building itself, and proactive matching and interaction between power load, renewable energy generation, and grid status. This transforms the building from a traditional single power load into an integrated "source-storage-load-flexibility" comprehensive energy node with power generation, regulation, and storage capabilities, supporting the stable operation of the new power system and promoting the low-carbon transformation of buildings. The multi-source heterogeneous data fusion panoramic monitoring method based on PV-Storage-DC-Flexible systems in this invention is specifically designed to accurately and in real-time perceive and analyze the panoramic operating status of this new integrated energy node in large-scale, distributed application scenarios, supporting its internal collaborative optimization and efficient interaction with the external power grid.

[0022] Therefore, this invention provides a panoramic monitoring method based on photovoltaic-storage-direct-drive-flexible (PV-SHU) architecture, applicable to a PV-SHU-direct-drive-flexible building panoramic monitoring system (hereinafter referred to as the "system"). This system can run on a distributed computing platform composed of edge computing nodes and a cloud data center via software services or a microservice architecture to complete panoramic data acquisition, processing, fusion, and result publishing of PV-SHU-direct-drive-flexible devices, energy and carbon emissions, environmental conditions, and grid interaction status within the building. Specifically, this system can be deployed in various hardware environments, including but not limited to: industrial gateways or embedded devices deployed on-site (as edge computing nodes), standard servers in data centers, private or public cloud virtual machine clusters, and containerized orchestration platforms (such as Kubernetes). This flexible deployment architecture allows the system to meet the low-latency, high-reliability requirements of distributed PV-SHU-direct-drive-flexible systems for data acquisition and real-time processing, while also adapting to the centralized, high-computing-power requirements of large-scale data aggregation, deep analysis, and model training.

[0023] like Figure 1 As shown, the method may include: Step S110: Configure multi-dimensional heterogeneous data sources covering photovoltaic-storage-direct-flexible equipment, building body, environment and power grid side, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer.

[0024] Among them, multidimensional heterogeneous data sources refer to a collection of various data sources with different types, formats, communication protocols, and acquisition frequencies, which together constitute a comprehensive description of the monitored object; edge data acquisition agents refer to integrated hardware and software modules deployed physically close to the data source (such as the equipment site), which are responsible for on-site acquisition, preliminary processing, and forwarding of raw data; distributed panoramic perception data acquisition layer refers to a unified data acquisition system composed of multiple geographically or logically dispersed edge data acquisition agents interconnected through a network, which aims to achieve collaborative acquisition of wide-area, multi-source data.

[0025] Specifically, the system can establish stable connection channels with various data sources by calling the Software Development Kit (SDK) provided by the equipment manufacturer or standard industrial communication protocols (such as Modbus TCP, MQTT), and configure the acquisition tasks and forwarding rules of the edge data acquisition agent. Modbus TCP (Modbus over TCP / IP) is an industrial communication protocol based on TCP / IP networks, used for deterministic, master-slave data exchange between industrial automation equipment (such as PLCs, sensors, and actuators). MQTT (Message Queuing Telemetry Transport) is a lightweight message transmission protocol based on a publish / subscribe model, designed for low-bandwidth, high-latency, or unstable network environments, and widely used for asynchronous message communication between IoT devices and servers.

[0026] For example, the system can deploy an edge data acquisition agent in the power distribution room of an office building. The agent reads data such as power generation and DC voltage from a photovoltaic inverter in real time via the Modbus TCP protocol or from a grid-connected photovoltaic inverter. It also obtains environmental data from indoor temperature and humidity sensors via the MQTT protocol. At the same time, it periodically obtains building space zoning information through the application programming interface (API) of the Building Information Modeling (BIM) platform, and sends this data to the cloud data center after unified encapsulation.

[0027] Step S120: Perform data cleaning, scale normalization, timestamp synchronization and spatial registration operations sequentially on the monitoring data collected by the distributed panoramic perception data acquisition layer to generate a standardized data set that is spatiotemporally aligned.

[0028] Data cleaning refers to the process of identifying and processing outliers, missing values, and noise in the original data; scale normalization refers to mapping data with different physical units and value ranges to a unified standard interval (e.g., scale normalization). or The process of cleaning, normalization, timestamp synchronization, and spatial registration is to eliminate the influence of dimensions; timestamp synchronization refers to the process of uniformly aligning data acquired at different collection frequencies or from different clock sources to the same time reference sequence through interpolation, resampling, and other methods; spatial registration refers to the process of establishing a unified spatial coordinate system and reference system for data from different data sources that have spatial attributes (such as equipment location, room partition); the standardized data set after spatiotemporal alignment refers to a high-quality data set that has consistent time dimension, spatial dimension, and data scale after completing the cleaning, normalization, timestamp synchronization, and spatial registration processes and is stored in a structured database or data warehouse.

[0029] Specifically, the system can receive raw data streams from various edge nodes through a data preprocessing service deployed in the cloud. This service first uses statistical distribution and isolated forest algorithms to detect and filter outliers such as transient jumps, and employs linear interpolation to fill in missing data caused by communication interruptions. Then, it performs max-min normalization on numerical data and one-hot encoding on categorical data. Next, it performs linear resampling on all time-series data at 1-minute intervals to achieve timestamp alignment. Finally, it uniformly converts the device location coordinates obtained from the building information model to the same building plane rectangular coordinate system as the IoT sensor network.

[0030] For example, the system processes temperature data from two data sources, A and B. Source A collects data every 5 seconds, and source B collects data every 1 minute. After the timestamp synchronization operation, the two data sequences are uniformly resampled into a time series with one data point per minute. The original value range of source A is... Celsius, after being normalized to its maximum and minimum values, is mapped to... Intervals. Ultimately, the two sequences, along with similarly processed data from other data sources, are stored together in a time-series database using unified timestamps and spatial location labels, forming the aforementioned standardized data set.

[0031] Step S130: Build and execute a collaborative data fusion computing pipeline from edge to cloud.

[0032] The pipeline includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge; a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud; and a collaborative fusion stage that performs coupled calculations on system operating status and energy and carbon indicators. The shallow fusion stage refers to the initial integration and denoising of similar or strongly correlated real-time, high-frequency data streams at edge computing nodes close to the data source. The deep fusion stage refers to the process of using deep learning models to perform deep feature association and extraction on the shallowly fused data and multimodal information such as static model parameters in the cloud, which has powerful computing capabilities. The collaborative fusion stage refers to the process of calculating comprehensive and decision-making indicators such as energy efficiency, carbon emissions, and flexible adjustment potential based on the features extracted by deep fusion, combined with business rules and physical models, in the cloud.

[0033] Specifically, the system can achieve the shallow fusion stage by deploying lightweight fusion computing services at the edge gateway. For example, it can perform weighted averaging of current and voltage data from multiple photovoltaic modules at the same time to estimate the real-time output power of the entire photovoltaic string and filter out sensor noise. In the cloud, the system can deploy a deep learning model based on attention mechanisms and graph neural networks (GNNs) to receive fused time-series power data from the edge, equipment topology data from the building information model, and weather forecast data. Through model training, it learns the complex nonlinear relationships between these data and outputs a feature vector representing the overall operational health and energy efficiency potential of the system. Finally, in the collaborative fusion computing module, the system can use this feature vector, combined with real-time grid carbon emission factors and time-of-use pricing, to calculate the building's current carbon intensity and potential economic benefits from participating in demand response.

[0034] For example, in a data center scenario, the system's edge nodes perform moving average filtering on real-time power consumption data (once per second) from 10 server racks, outputting a second-level power consumption curve for the entire IT load to the cloud. A cloud-based deep learning model, combining this power consumption curve with data such as outdoor temperature and humidity, and chiller unit operating status, analyzes and concludes that the cooling system efficiency is declining. The collaborative fusion module further calculates, based on this trend and the current real-time carbon emission factor of the power grid (e.g., 0.5 kg CO2 per kilowatt-hour), that maintaining the current state would result in approximately 50 kg of additional CO2 emissions in the next hour, and provides optimization suggestions for prioritizing adjustments to the chiller unit's settings.

[0035] Step S140: Publish the fusion features and indicators output from the collaborative fusion phase through a standardized application programming interface.

[0036] Standardized application programming interfaces (APIs) refer to communication interfaces between software components that conform to specific industry standards or general design specifications (such as RESTful architecture and gRPC framework), used to achieve standardized calls to data and functions between different systems or services. Fusion characteristics and indicators refer to structured data generated after the aforementioned collaborative fusion stage, which can directly characterize the system state and support management decisions, such as system energy efficiency assessment values, real-time carbon emission intensity, and adjustable capacity of flexible loads. RESTful (Representational State Transfer architectural style) architecture refers to a software architectural style based on the HTTP protocol and following specific design principles. gRPC (Google Remote Procedure Call framework) framework refers to a high-performance, open-source, and general-purpose remote procedure call (RPC) framework developed by Google.

[0037] Specifically, the system can design the application programming interface based on the Representational State Transfer (REST) ​​architecture, encapsulate the fused features and metrics into data objects in JSON (JavaScript Object Notation) or Protocol Buffers format, and provide a GET method or subscription push interface based on the Hypertext Transfer Security Protocol (HTTPS).

[0038] For example, the system encapsulates the "predicted carbon emission increase of 50 kg in the next hour" and the "chiller unit adjustment suggestion" calculated in step S130 into a JSON object. This object is published through a RESTful application programming interface, and its Uniform Resource Locator (URL) is https: / / XXXXX / XX / XXXXX / XXXXX. The building energy management system or regional power grid dispatching platform can periodically obtain this data by sending HTTPS GET requests to this address, and automatically or assistedly generate operation and dispatching instructions based on it.

[0039] In another embodiment, such as Figure 2This demonstration showcases the implementation architecture of a "panoramic monitoring system based on multi-source heterogeneous data fusion of photovoltaic, energy storage, DC, and flexible power distribution" technology applied to a large commercial complex. The complex has a building area of ​​approximately 200,000 square meters and integrates distributed photovoltaic systems, energy storage systems, DC power distribution networks, and flexible air conditioning and lighting loads. In this embodiment, the multi-dimensional data acquisition layer is specifically deployed as follows: edge acquisition terminals are deployed in the rooftop photovoltaic area and underground energy storage power station, aggregating real-time operating parameters of photovoltaic modules, energy storage batteries, and DC power distribution cabinets via industrial Ethernet and 5G mobile communication networks; IoT sensor nodes are deployed on each floor of the building to collect temperature, humidity, illuminance, and sub-item energy consumption data; the equipment spatial topology in the building information model is synchronized through an application programming interface (API), and real-time electricity prices and regional carbon emission factor platform data published by the power grid company are accessed. The data preprocessing and standardization layer performs real-time cleaning and normalization of the multi-source data at the second to minute level, and completes the spatiotemporal alignment of all time-series data with a 1-minute benchmark, forming a unified and standardized data stream. In the core layer of multi-source heterogeneous data fusion, the system first performs adaptive Kalman filtering on the second-level current and voltage data of each branch of the photovoltaic array on the edge server close to the data source to achieve shallow fusion and output a stable and denoised estimate of the total power of the array. Subsequently, the fused power sequence, building cooling load data, outdoor meteorological data, and equipment topology are uploaded to the cloud and input into a deep learning model based on Transformer and graph neural network for deep fusion. The model discovers the cross-modal correlation features between "afternoon afternoon sun causing a surge in local cooling load" and "PV output fluctuations due to cloud cover". Finally, based on these correlation features, the collaborative fusion module, combined with real-time electricity prices and carbon emission factors, calculates that if the current operation is maintained in the next 15 minutes, the carbon emission intensity will increase by 5%, and generates an optimization strategy to reduce the non-critical lighting load in the western area by 20% to balance the photovoltaic fluctuations. Finally, at the application service layer, the strategy triggers the photovoltaic-storage-direct-flexible coordinated scheduling module, issuing flexible adjustment commands to the building automation system. Simultaneously, updated carbon emission data is pushed to the building energy and carbon management platform for real-time carbon accounting dashboards. Aggregated building net load adjustment potential data is uploaded to the urban power grid dispatch center, supporting source-grid-load-storage cluster control decisions. This embodiment demonstrates how the panoramic monitoring technology extracts decision-making information from dispersed sensing data and drives multi-level coordinated optimization from individual buildings to the regional power grid.

[0040] In another embodiment, Table 1 below shows the measured data analysis of the end-to-end real-time performance guarantee capability of the control link of this panoramic monitoring scheme in a scenario where a "photovoltaic-storage-DC-flexible" aggregate participates in primary frequency regulation of the power grid. This embodiment selects a virtual power plant (VPP) composed of photovoltaic and energy storage resources from five commercial buildings as the test object. In this scenario, the power grid dispatching agency requires the aggregate to complete power adjustment within 2 seconds after sensing a frequency deviation, which places stringent requirements on the latency of the monitoring and control link. This scheme optimizes each link to ensure that the end-to-end full-link latency from data acquisition to control command issuance is ≤1.7 seconds, fully meeting the real-time control requirements. Table 1 details the measured maximum latency and implementation method for each link: This embodiment demonstrates that by adopting a collaborative architecture of "edge intelligent acquisition + 5G slicing transmission + local rapid fusion," this solution successfully reduces the end-to-end latency of the monitoring and control link to less than 1.7 seconds, providing reliable technical support for applications with stringent real-time requirements, such as the participation of "photovoltaic-storage-DC-flexible" resources in rapid grid frequency regulation. The deep fusion and collaborative fusion results performed in the cloud (for long-term strategy optimization, latency ≤ 5 seconds) are dynamically updated using edge-side fusion algorithm parameters (such as weights and droop curves), continuously improving the accuracy and economy of control without interfering with the real-time control link.

[0041] Therefore, according to the above implementation method, the system achieves its goals through four core steps: panoramic perception construction, data standardization generation, collaborative fusion computing, and standardized result release. Specifically, it configures multi-dimensional heterogeneous data sources covering photovoltaic-storage-direct-source and flexible photovoltaic systems, building structures, the environment, and the power grid, and deploys edge data acquisition agents to construct a distributed panoramic perception data acquisition layer. This provides a comprehensive, traceable monitoring data foundation covering the entire spectrum of "source-grid-load-storage-loop" for subsequent fusion computing. The monitoring data acquired by this acquisition layer undergoes data cleaning, scale normalization, timestamp synchronization, and spatial registration operations to generate a spatiotemporally aligned standardized data set. This eliminates differences in units, sampling frequencies, and spatial reference systems among the multi-source heterogeneous data, providing a unified and standardized input for high-quality data fusion. Finally, it constructs and executes collaborative data fusion from the edge to the cloud. The fusion computing pipeline achieves shallow fusion by performing adaptive filtering on real-time data streams at the edge to ensure real-time processing. It then performs deep learning feature extraction on cross-modal data in the cloud to achieve deep fusion and uncover complex correlations. Finally, it performs coupled calculations on system operating status and energy and carbon indicators to achieve collaborative fusion and generate fusion indicators that can directly serve optimization decisions. This forms a progressive fusion capability from real-time processing to deep analysis and then to application-oriented approaches. The fusion features and indicators output from the collaborative fusion stage are published through standardized application programming interfaces (APIs), achieving loose coupling and standardized integration between monitoring results and upper-level scheduling, control, and optimization applications.

[0042] Specifically, in the technical solution of this embodiment, to address the problem of insufficient fusion accuracy caused by incomplete data source coverage and spatiotemporal misalignment as described in the background technology, a distributed panoramic perception data acquisition layer is constructed and strict spatiotemporal alignment operations are performed on multi-source data. This ensures the completeness and consistency of input data from the data source and preprocessing levels, laying a solid foundation for improving fusion accuracy. To address the problem of difficulty in balancing real-time performance and accuracy due to the lack of a fusion architecture adapted to the characteristics of distributed, highly volatile data, a collaborative computing pipeline consisting of shallow edge-side fusion, deep cloud-side fusion, and collaborative fusion is designed. Edge computing is used to ensure the real-time data... The initial low-latency processing of the data stream leverages cloud computing power to support deep correlation mining and complex index coupling calculations across modal data, thereby achieving a balance between real-time data processing and analytical depth at the architectural level. Addressing the issue of low utilization efficiency and ineffective support for collaborative scheduling and carbon efficiency optimization caused by the disconnect between monitoring data and advanced applications, this invention directly couples the calculations of energy and carbon indicators such as operating status, carbon emissions, and the value of flexible regulation during the collaborative fusion stage. The final results are then published through a standardized application programming interface, enabling the output fusion features and indicators to be directly and efficiently invoked by optimization models for photovoltaic-storage-DC-flexible collaborative scheduling, building energy and carbon management, and source-grid-load-storage cluster control. Therefore, the technical solution of this invention, through the systematic synergy of the above steps, solves the technical problems of insufficient accuracy, real-time performance, and matching degree between the existing technology and the requirements of system collaborative scheduling and carbon efficiency optimization when facing the large-scale application of photovoltaic-storage-DC-flexible systems. This improves the accuracy, responsiveness, and support capability for low-carbon operation decisions in buildings and new power systems.

[0043] In some embodiments, the system can perform scale normalization processing on multi-source heterogeneous monitoring data using the following formula (a): (a) Formula (a) is the Z-score standardization formula, used to eliminate the influence of different physical dimensions on subsequent fusion calculations. x represents the original data value, such as photovoltaic power generation at a certain moment (unit: kilowatt, kW), room temperature (unit: degree Celsius, ...). The system uses either the power output of distributed photovoltaic (PV) or the state of charge (a dimensionless percentage); μ represents the arithmetic mean of the data within a specific time window; σ represents the standard deviation of the data within the same time window, characterizing the dispersion of the data; and x' represents the standardized dimensionless value. In this embodiment, to adapt to the large fluctuations in distributed PV output, the system uses a 1-hour sliding time window to dynamically calculate μ and σ for time-series data (such as power and temperature); for static or quasi-static data (such as dimensional parameters in a building information model), a fixed statistical mean is used as μ, and typical process error is used as σ. Through this transformation, data from different sources and with different dimensions are mapped to the same numerical scale with a mean of 0 and a standard deviation of 1.

[0044] Next, in this embodiment, the system can also perform adaptive Kalman filter fusion on real-time sensing data (such as distributed photovoltaic and energy storage power) using the following formula (b): (b) Equation (b) is the core equation for measurement update in Kalman filtering, which is used to achieve complementary optimization and state estimation of real-time data of the same dimension. The optimal estimate of the system state (such as the actual power) at time k; This represents the prior prediction of the state at time k based on the state at time k-1. H represents the actual observed value (sensor reading) at time k; H is the observation matrix, which maps the state space to the observation space. The Kalman gain matrix determines the observed values. For the final estimate The corrected weights are adjusted. In this embodiment, in order to suppress high-frequency data jitter, the system dynamically adjusts the process noise covariance matrix Q (range 0.01~0.05) and the observation noise covariance matrix R (range 0.1~0.3), so that the algorithm can adapt to the fluctuation characteristics of photovoltaic and energy storage data and achieve efficient noise reduction and state tracking.

[0045] Furthermore, in this embodiment, the system can also perform a reliability-weighted average fusion of static ledger data (such as building information model equipment parameters and energy efficiency ratings) using formula (c): (c) Formula (c) is a variance-based inverse weighting formula used to assign fusion weights to static data sources with different levels of reliability. This represents the normalized weight assigned to the i-th data source; The standard deviation of the i-th data source is used to quantify the uncertainty or noise level of that data source; the denominator is summed over all j data sources participating in the fusion to ensure the total weight. In this embodiment, for data from different archives or building information models of different accuracy levels, the system sets the accuracy level based on its labeled accuracy level or historical verification error. The lower the uncertainty ( Smaller data sources receive higher weight in the fusion process. This improves the reliability of the fusion results.

[0046] Furthermore, in this embodiment, the system can also perform supervised training of the multimodal deep learning model in the deep fusion inference engine using the composite loss function shown in formula (d): , (d) Formula (d) defines the overall objective of model training. L represents the total loss. Represents the mean squared error loss, used to constrain the model's predicted values. With real labels The degree of similarity (such as historical energy consumption and measured carbon emissions); The contrastive learning auxiliary loss is used to bridge the gap between related multimodal features (such as time-series power features and device topology features) in the representation space, promoting the model's learning of cross-modal correlations. λ is the contrastive loss weight coefficient, determined through cross-validation, with a recommended value range of 0.01 to 0.1. In this embodiment, the training data is divided into training, validation, and test sets in a 7:2:1 ratio. Training stops early when the validation set loss does not decrease for 20 consecutive rounds, and the model is fine-tuned online every quarter with new data to maintain adaptability to changes in building operation status.

[0047] Furthermore, in this embodiment, the system can also calculate the real-time carbon intensity of building electricity consumption during the collaborative fusion phase using formula (e): (e) Formula (e) is the core formula for energy-carbon coupling calculation. The building's total carbon intensity at time t is expressed in kilograms of carbon dioxide equivalent per kilowatt-hour (kWh). ); This represents the amount of electricity purchased from the grid at time t (unit: kWh). This represents the amount of electricity purchased from the grid at time t (unit: kWh). Represents the real-time carbon emission factor of the power grid at time t (unit: ), dynamically obtained from external platforms; The amount of electricity generated by photovoltaic power generation and consumed locally at time t (unit: kWh). Carbon emission factor representing the entire life cycle of photovoltaic power generation (unit: ), which are fixed parameters; This represents the total electricity consumption of the building at time t, satisfying... This formula quantifies the net carbon emissions per unit of electricity consumption, directly reflecting the level of low-carbon energy use in buildings.

[0048] Finally, in this embodiment, the system can also evaluate the carbon reduction and economic benefits of flexible load regulation using formulas (f) and (g), and construct the regulation value matrix shown in formula (h): (f) (g) ] ;(h) Formula (f) calculates the carbon emission reduction achieved by adjusting device i at time t. (unit: Formula (g) calculates the corresponding electricity cost savings. (Unit: Yuan) The real-time carbon intensity is calculated using formula (e); The real-time time-of-use electricity price at time t (unit: yuan / kWh); The power adjustment range of device i (unit: kW); The duration of the adjustment action (in hours) is used. Formula (h) combines the carbon reduction and electricity savings of each adjustable device at each moment into a two-dimensional vector, forming the "adjustment value matrix". Based on this matrix, the scheduling platform can prioritize the use of equipment with high adjustment value, achieving optimal decision-making under the dual objectives of economic efficiency and low carbon emissions.

[0049] Therefore, by combining the above formulas (a) to (h), the system can construct a complete and rigorous mathematical calculation system, thereby realizing the standardized preprocessing of monitoring data, the reliable fusion of multi-source heterogeneous data, the accurate training and optimization of deep learning models, and the real-time coupled evaluation of the operating energy and carbon status. Ultimately, it quantifies the multi-faceted value of flexible regulation, providing a quantitative data foundation and evaluation basis that can be directly used for optimization decision-making for the coordinated scheduling of photovoltaic-storage-direct-flexible systems and the optimization of building carbon efficiency.

[0050] In some embodiments, a multi-dimensional heterogeneous data source covering photovoltaic-storage-direct-drive-flexible equipment, building structure, environment, and power grid is configured, and an edge data acquisition agent is deployed to construct a distributed panoramic perception data acquisition layer, including: Configure heterogeneous data source access endpoints covering the operating status of photovoltaic, energy storage, direct current and flexible photovoltaic equipment, building IoT sensor nodes, building information model, power grid interaction interface and third-party environmental data platform.

[0051] Among them, heterogeneous data source access endpoints refer to the software interfaces or communication addresses configured for each specific type or protocol of data source, which are used to establish a dedicated connection channel between the system and the data source.

[0052] Specifically, the system can create the access endpoint by filling in the necessary connection parameters for each type of data source in the configuration management interface. These parameters may include Internet Protocol (IP) address, port number, Application Programming Interface (API) key, authentication token, data query statement, etc. For example, the system configures an access endpoint based on the Modbus TCP protocol for a photovoltaic inverter, with a target IP address of 192.168.1.100 and port 502; and configures an HTTPS-based RESTful API access endpoint for the building information modeling platform, with a Uniform Resource Locator (URL) of https: / / XXXXXXX / XX / XX / XXXXXX and a valid OAuth 2.0 access token.

[0053] At the physical proximity of the optical storage direct current and flexible equipment and the building IoT sensor network, an edge data acquisition agent is deployed. The edge data acquisition agent is used to perform data extraction, transformation and loading processes on the raw data stream originating from the local area and temporarily store it in the edge data lake.

[0054] Among them, the data extraction, transformation and loading process refers to the complete data processing process of obtaining raw data from the data source, converting it into a unified format and structure defined within the system, and loading it into the target storage area; the edge data lake refers to the storage area established on the edge computing node for storing various types of raw, unprocessed or only preliminarily processed data, and its characteristic is that it supports the coexistence of multiple data formats.

[0055] Specifically, after startup, the edge data acquisition agent periodically "extracts" raw data from the data source through the access endpoint according to predefined acquisition tasks. The agent then "converts" the data into a unified JavaScript Object Notation (JSON) format and adds metadata tags such as timestamps and data source identifiers. Finally, the agent "loads" the processed data into a local embedded database or file for temporary storage. For example, an edge data acquisition agent deployed in a substation "extracts" current, voltage, and power data from a local smart meter every 5 seconds. The agent "converts" the raw meter messages into JSON records such as {"timestamp":"2023-10-27T10:00:00Z","device_id":"meter_01","power_kW":150.5} and "loads" them into a local SQLite database. This database is the edge-side data lake, which also stores heterogeneous data from other local sensors (such as temperature and humidity).

[0056] Define differentiated data transmission service quality levels and routing strategies for the communication link between edge data acquisition agents and cloud data centers.

[0057] Among them, the data transmission service quality level refers to the performance guarantee level set for different types of data streams regarding transmission bandwidth, latency, jitter, and reliability; the routing policy refers to the selection rules that determine the network path that data takes from the edge to the cloud.

[0058] Specifically, the system can achieve differentiated transmission by configuring Software-Defined Networking (SDN) policies or the Quality of Service (QoS) level of message middleware (such as MQTT Broker). High real-time data (such as fault alarms) is assigned the highest QoS level (such as QoS2) to ensure delivery and low latency; routine monitoring data uses a medium QoS level (such as QoS1); while non-real-time data such as log files uses a best-effort QoS level (such as QoS0). For example, the system defines the following strategy: second-level power data of photovoltaic inverters (high real-time) is transmitted through dedicated 4G (fourth-generation mobile communication technology) network slices, with an end-to-end latency requirement of less than 100 milliseconds; building indoor temperature data (routine monitoring) is transmitted through the enterprise wireless LAN (Wi-Fi), with a latency requirement of less than 5 seconds; and equipment operation log files (non-real-time) are transmitted in batches during idle periods of wired network bandwidth at night.

[0059] In the cloud data center, maintain the metadata registry of heterogeneous data source access endpoints and edge data acquisition agents to complete the logical integration of the distributed panoramic perception data acquisition layer.

[0060] The metadata registry is a directory service that centrally stores and manages descriptive information (i.e., metadata) of all data sources and acquisition agents, used to realize resource discovery, management and status monitoring.

[0061] Specifically, the system can deploy a registry service in the cloud. Each heterogeneous data source access endpoint, upon creation, and each edge data acquisition agent, upon startup, must register with this registry, reporting its unique identifier, type, network address, health status, and the data pattern being collected. For example, the cloud metadata registry records an access endpoint information entry: {“id”:“endpoint_pv_01”,“type”:“modbus_tcp”,“address”:“192.168.1.100:502”,“status”:“online”,“data_schema”:“pv_power,dc_voltage”}. Simultaneously, a data acquisition agent information entry was recorded: {“id”:“agent_room_201”,“location”:“BuildingA / Floor2 / Room201”,“endpoints”:[“endpoint_pv_01”,“endpoint_temp_01”],“last_heartbeat”:“2023-10-27T10:00:30Z”}. By querying this registry entry, the system can gain a global understanding of the data acquisition layer's topology and real-time status.

[0062] Therefore, according to the above implementation method, the system can achieve unified, manageable and reliable acquisition of wide-area, distributed and heterogeneous multi-source data, laying a solid data foundation for subsequent data processing and fusion, and effectively addressing the monitoring challenges brought about by the dispersed deployment of distributed optical storage direct current and flexible systems.

[0063] In some embodiments, the monitoring data collected by the distributed panoramic perception data acquisition layer is sequentially subjected to data cleaning, scale normalization, timestamp synchronization, and spatial registration operations to generate a spatiotemporally aligned standardized data set, including: An instantiated data cleaning computation graph is used to sequentially perform anomaly detection and filtering based on statistical distribution and isolated forest, as well as missing value imputation based on interpolation algorithm on the input monitoring data stream.

[0064] In this context, a data cleaning computation graph refers to a directed acyclic computational structure that defines the various steps of data cleaning (such as inputs, multiple processing operators, and outputs) and their dependencies, and is used to guide the execution of the cleaning process.

[0065] Specifically, the system can construct a computation graph in memory by calling the application programming interface (API) of a data processing framework (such as Apache Spark or Apache Flink). This computation graph first receives the raw data stream and then executes two cleaning operators in parallel: one operator is based on 3 sigma (…). One criterion identifies and removes numerical points that significantly deviate from the historical statistical distribution; another operator, based on the Isolation Forest algorithm, detects and filters out anomalies caused by transient equipment failures that are clearly isolated from the overall data pattern. For data points marked as anomalies or missing, the computational graph calls a fill operator based on linear interpolation to estimate and fill the missing data points using valid values ​​before and after the data point. For example, the system processes a second-level voltage data stream from an energy storage battery management system. The data cleaning computational graph detects that at the timestamp 2023-10-27T14:30:05, the voltage value jumps instantaneously from the normal 750 volts (V) to 1200V, exceeding the value calculated based on historical data. The range (e.g., 700V to 800V) is defined by the point " The "Criterion" operator was flagged as an anomaly. Additionally, at timestamp 2023-10-27T14:30:10, data was missing due to a brief communication interruption. The "Linear Interpolation" operator used 749.8V at 2023-10-27T14:30:04 and 750.1V at 2023-10-27T14:30:12 to calculate a filler value of approximately 750.0V for 2023-10-27T14:30:10.

[0066] Perform feature scaling or standardization transformations on the cleaned data sequence to eliminate the influence of different physical dimensions on subsequent calculations.

[0067] Feature scaling and normalization are two mathematical methods for transforming data to a uniform scale range. Feature scaling (such as max-min normalization) linearly maps data to a specific interval (e.g., ...). Standardization (such as Z-score standardization) transforms data into a distribution with a mean of 0 and a standard deviation of 1.

[0068] Specifically, the system can maintain a mapping configuration table between data patterns and transformation methods. For numerical features, the system selects a transformation method based on the configuration: if the data distribution is bounded and has no extreme values, maximum-minimum normalization can be used, with the formula as follows: If the data distribution may contain tails or a more stable scale is required, Z-score standardization can be used, with the formula being: ,in The mean, The standard deviation is given. For categorical features, one-hot encoding is used for transformation. For example, the system needs to process photovoltaic DC power (unit: kilowatts, kW) and indoor temperature (unit: degrees Celsius). Two sequences. Historical maximum photovoltaic power ( The minimum value is 100 kilowatts. The power is 0 kW. At a certain moment, the power is 60 kW, and the value after normalization to the maximum and minimum values ​​is... The mean of the indoor temperature series ( ) is 25 Standard deviation ( ) is 3 The temperature at a certain moment was 28 degrees Celsius. The value after Z-score standardization is After transformation, data from two different units are converted to a comparable numerical scale.

[0069] The spatiotemporal alignment service is invoked to perform upsampling or downsampling operations on heterogeneous time-series data streams at a preset reference sampling frequency, and coordinate system one and spatial index construction are performed on static or quasi-static model data with spatial attributes to achieve consistent alignment of multi-source data on timestamps and spatial reference systems.

[0070] Among them, the spatiotemporal alignment service is a background service specifically responsible for coordinating the consistency of the time and spatial dimensions of data; upsampling refers to the process of converting low-frequency time series data into higher-frequency data through interpolation and other methods; downsampling refers to the process of converting high-frequency time series data into lower-frequency data through aggregation (such as averaging, summing) and other methods.

[0071] Specifically, the system can preset a baseline sampling frequency (e.g., 1 minute). The spatiotemporal alignment service receives time-series data streams from different data sources. If the original frequency of a data stream is higher than the baseline frequency (e.g., at the second level), the service performs a downsampling operation on it, for example, calculating the average of all second-level data within each minute as the representative value for that minute. If the original frequency is lower than the baseline frequency (e.g., at the 5-minute level), the service performs an upsampling operation on it, for example, using forward padding to fill the missing values ​​of the current minute with the previous valid value. For device coordinates from the Building Information Model (BIM) (which may be based on the project coordinate system), the service calls a coordinate transformation library to convert them to a building global Cartesian coordinate system consistent with the IoT sensor network (e.g., a coordinate system with the lower left corner of the building as the origin), and builds an R-tree spatial index for all data points with spatial locations to achieve fast spatial range lookup. For example, light sensor data is reported once per second, while air conditioning unit power data is reported once every 5 minutes. The spatiotemporal alignment service uses a 1-minute baseline to calculate the average value per minute for illumination data (downsampling). For air conditioning power data, assuming data is available at 10:00, data points from 10:01 to 10:04 are filled with the 10:00 value (upsampling). Simultaneously, the service will specify the location coordinates of "Air Conditioning Unit 01, Area A, 3rd Floor" as described in the BIM. (Unit: millimeters), converted to a unified coordinate system (Unit: meters), and create an index.

[0072] Data units that have undergone feature scaling and spatiotemporal alignment are partitioned and cataloged according to their data patterns and business domains, and then persistently stored as a standardized data set that has undergone spatiotemporal alignment.

[0073] Among them, a data unit refers to a single data record formed after all the aforementioned processing, which is accompanied by a unified timestamp, spatial label and standardized value; partitioning and cataloging refer to dividing the data into different physical or logical storage areas according to certain attributes of the data (such as time, source and type) and establishing a directory structure that is easy to retrieve.

[0074] Specifically, the system can write the processed data to a time-series database (such as InfluxDB or TDengine) or a big data platform (such as Hive). During storage, the system partitions the data according to a hierarchy of "data mode / business domain / time". For example, it creates independent tables or measurements for different business domains such as "photovoltaic power generation," "environmental monitoring," and "equipment operation." Within each table, the data can be further partitioned according to "year-month-day" or even "hour." Simultaneously, the system updates a global metadata directory, recording which time ranges and data types are stored in each partition. For example, the processed data for 10:00 AM on October 27, 2023, includes one photovoltaic power record {time: "2023-10-27T10:00:00Z", location: "roof_south", value: 0.6} and one temperature record {time: "2023-10-27T10:00:00Z", location: "room_301", value: 1.0}. The system writes these records to tables named pv_power and env_temperature, respectively. The pv_power table is partitioned by date, and this record is stored in the 2023 / 10 / 27 partition. The metadata directory records that the pv_power table contains data from 00:00:00 to 23:59:59 in partition 2023 / 10 / 27.

[0075] Therefore, according to the above implementation method, the system can transform raw monitoring data from different sources, with different formats and different spatiotemporal references into a high-quality, correlated, and directly usable standardized data asset for advanced analysis and calculation, providing an accurate and consistent input basis for subsequent data fusion and calculation.

[0076] In some embodiments, the step of building and executing a collaborative data fusion computing pipeline from the edge to the cloud includes: A shallow fusion service is instantiated on the edge computing node. The shallow fusion service is used to perform adaptive filtering and fusion calculations on the real-time high-dimensional data stream from the distributed panoramic perception data acquisition layer and output low-dimensional fusion temporal features.

[0077] Shallow fusion service refers to a software service deployed on edge computing nodes that focuses on real-time, lightweight integration and denoising of raw high-dimensional sensor data; adaptive filtering fusion computation refers to a real-time estimation algorithm that can dynamically adjust filtering parameters according to the noise statistical characteristics of the input data stream, used to extract signal estimates that are closer to the real state from noisy observation data.

[0078] Specifically, the system can instantiate the shallow fusion service by deploying a lightweight filtering algorithm library (e.g., a library implementing Kalman filtering or recursive least squares) in a container (such as a Docker container) on the edge node. This service receives second-level raw data streams from multiple similar sensors (e.g., current sensors for multiple photovoltaic modules), dynamically calculates the optimal fusion weights based on the historical covariance information of each data stream, performs weighted fusion of the multiple data streams, and corrects the state estimation in real time, ultimately outputting a fused data sequence with a frequency still on the second level that represents the overall state of the sensor group. For example, the shallow fusion service fuses the real-time current (unit: amperes, A) of three modules in the same photovoltaic array. At time t, the original observation value is... The service assesses the noise covariance matrix of each sensor observation at the current moment based on historical data. The fused array current estimate is obtained through adaptive weighted fusion calculation (e.g., based on minimum variance unbiased estimation). Amperes, and continuously output this fused sequence at a frequency of seconds.

[0079] Deploy a deep fusion inference engine in a cloud data center. The deep fusion inference engine receives low-dimensional fusion time-series features and static or quasi-static model parameter data from a standardized dataset. It then performs feature extraction and association learning on the heterogeneous data stream through a multimodal deep learning model to generate a system-level deep fusion feature vector.

[0080] Among them, the deep fusion inference engine refers to a software system deployed in the cloud that integrates and runs complex deep learning models to mine deep correlations across modal data; the multimodal deep learning model refers to a neural network architecture that can simultaneously process and correlate inputs from different data modalities (such as time series, images, graphs, and text); and the deep fusion feature vector refers to a high-dimensional, dense real-number vector extracted and output by the deep learning model, which compactly encodes the deep operational status and correlation information of the system across multiple modalities.

[0081] Specifically, the system can deploy a model service based on the PyTorch or TensorFlow framework in the cloud. This service loads a pre-trained multimodal model, which may include a Long Short-Term Memory (LSTM) branch to process the fused temporal features, a Graph Convolutional Network (GCN) branch to process the device topology graph extracted from the Building Information Model (BIM), and an attention mechanism module to fuse temporal and graph structure features. The model receives the fused temporal features of the photovoltaic-storage system power and the system device connection graph from the past hour, and through forward propagation, outputs a 256-dimensional real-valued vector as the deep fusion feature vector. For example, the deep fusion inference engine receives input: minute-level fused temporal sequences (60 points each) of the past 60 minutes of "total power of the photovoltaic array" and "total building load," and an adjacency matrix describing the connection relationships of "photovoltaic array-energy storage-critical load." The LSTM branch in the model learns the temporal patterns of the power sequence, the GCN branch learns the energy transfer influence relationships between devices, and the attention mechanism module identifies the cross-modal association that "the correlation between energy storage discharge behavior and specific loads increases when photovoltaic output decreases in the afternoon." Finally, the model outputs a 256-dimensional feature vector F, where high activation values ​​in some dimensions may correspond to abstract system state combinations such as "sufficient photovoltaic output," "energy storage in a charging state," and "low grid interaction power."

[0082] Deploy a collaborative fusion computing module in the cloud data center. Based on the deep fusion feature vector, the collaborative fusion computing module performs coupled calculations of operating status, carbon emission intensity and flexible adjustment value to generate multi-dimensional collaborative fusion indicators.

[0083] Among them, the collaborative fusion computing module refers to a software module that transforms the abstract features extracted by deep learning into specific and decision-making business indicators based on physical rules, business knowledge, and lightweight computing models; the coupled computing of operating status, carbon emission intensity, and flexible adjustment value refers to the calculation process that combines the features reflecting the physical operating status of the system with the carbon emission factor reflecting environmental impact and the electricity price signal reflecting economic value for comprehensive evaluation; and the multidimensional collaborative fusion index refers to a set of indicators composed of multiple specific values, which quantify the comprehensive performance of the system from different dimensions (such as energy efficiency, carbon emission, and economics).

[0084] Specifically, this module can contain a series of configurable computation rules and a lightweight feedforward neural network. The module receives the 256-dimensional deep fusion feature vector F, and first maps it to several basic state scalars, such as "system net power," through a fully connected layer network. (Unit: kilowatt, kW), "Comprehensive efficiency of photovoltaic-storage system" Then, it combines the real-time grid carbon emission factor (CEF, unit: kilograms of carbon dioxide per kilowatt-hour) obtained from external data sources. ) and real-time electricity price (Unit: Yuan per kilowatt-hour, ¥ / kWh), perform coupled calculations: carbon emission intensity (Unit: kilograms of carbon dioxide per hour) ), flexible adjustment value (Unit: Yuan per hour, ¥ / h). η, CI, and FV together constitute the multi-dimensional collaborative fusion index. For example, the module calculates the current system net power based on the feature vector F. kW (negative values ​​indicate power fed into the grid), efficiency The query revealed the current carbon emission factor of the power grid. Real-time electricity price =0.8¥ / kWh. Therefore, the current carbon emission intensity is calculated. (Negative values ​​indicate reduced carbon emissions), the value of flexible adjustment. (Negative values ​​indicate gains). The final set of metrics is: Net power: −15kW, System efficiency: 92%, Carbon intensity: −7.5 Adjusted value: −12¥ / h.

[0085] Multidimensional collaborative fusion indicators are serialized into a standardized data exchange format and published through a standardized application programming interface.

[0086] Serialization refers to the process of converting data structures or object states in memory into a standardized byte stream that can be stored or transmitted.

[0087] Specifically, the system can use a JSON serialization library (such as Jackson in Java or the json module in Python) to convert the multi-dimensional collaborative fusion indicator dictionary object into a JSON string. Then, this JSON string is returned as a response body to the caller via a RESTful application programming interface (API) based on the HTTP protocol, or published in a specified topic in a message queue (such as Kafka or RabbitMQ). For example, the collaborative fusion calculation module serializes the indicator dictionary {"net_power_kw":-15,"system_efficiency":0.92,"carbon_intensity_kgco2_per_h":-7.5,"flex_value_rmb_per_h":-12} into a JSON string: "{\"net_power_kw\":-15,\"system_efficiency\":0.92,\"carbon_intensity_kgco2_per_h\":-7.5,\"flex_value_rmb_per_h\":-12}. This string is sent to the “building / energy-metrics” topic on the internal message bus via a POST or PUT request, or provided to the building energy management system via the response of a GET request from the application programming interface endpoint https: / / XXXXXX / XX / XXXXXX.

[0088] Therefore, according to the above implementation method, the system can construct a complete and collaborative data value extraction chain from real-time preprocessing at the edge to deep feature learning and business indicator coupling calculation in the cloud, and finally transform the original and scattered monitoring data into standardized fused information products that can directly drive advanced application decisions.

[0089] In some embodiments, the fusion features and metrics output from the collaborative fusion phase are published through a standardized application programming interface, including: Define the data model and interface protocol specifications for the multi-dimensional collaborative integration indicators output during the collaborative integration phase.

[0090] Among them, the data schema refers to the formal description of the data structure, field names, data types, units, and allowed value ranges; the interface protocol specification refers to the standardized definition of the communication protocol, request / response format, authentication, error codes, and other interaction details of the application programming interface.

[0091] Specifically, the system can use a schema definition language (such as JSON Schema) to formally describe the data schema. Simultaneously, a detailed interface documentation should be written to define the interface protocol specifications, including using HTTP / 1.1 or HTTP / 2 protocols, adopting a Representational State Transmission (REST) ​​style, requiring callers to provide a valid API key for authentication in the request header, and specifying that the response body uses JSON format.

[0092] For example, the system defines the following JSON Schema data pattern for multi-dimensional collaborative fusion indicators: { "$schema":"http: / / XXXXXX / XXX / XXX#", "type": "object", "properties":{ "timestamp":{"type":"string","format":"date-time"}, “metrics”:{ "type": "object", "properties":{ "net_power_kw":{"type":"number","unit":"kilowatt"}, "system_efficiency":{"type":"number","minimum":0,"maximum":1}, "carbon_intensity_kgco2_per_h":{"type":"number","unit":"kilogram CO2per hour"}, "flex_value_rmb_per_h":{"type":"number","unit":"Renminbi per hour"} }, "required":["net_power_kw","system_efficiency","carbon_intensity_kgco2_per_h","flex_value_rmb_per_h"] } }, "required":["timestamp","metrics"] }

[0093] Multidimensional collaborative integration indicators are serialized into structured data exchange objects according to data patterns.

[0094] Serialization refers to the process of converting a data structure in memory into a standardized byte stream that conforms to the data pattern and can be transmitted. In this context, a structured data exchange object specifically refers to a JSON object string generated after serialization that conforms to the data pattern definition.

[0095] Specifically, after generating the metrics, the system calls a serialization library function to populate the metric data into a dictionary structure based on the constraints of the data schema, and then converts it into a JSON string. This process ensures that all field names, data types, and nesting structures fully conform to the predefined JSON schema. For example, the metric data in the system's memory might be a timestamp. Net power kilowatts, system efficiency 0.91, carbon emission intensity kilograms of carbon dioxide per hour, flexible adjustment value Yuan per hour. After serialization according to the data pattern, the following JSON string is generated: {"timestamp":"2023-10-27T14:30:00Z","metrics":{"net_power_kw":-15.2,"system_efficiency":0.91,"carbon_intensity_kgco2_per_h":-7.6,"flex_value_rmb_per_h":-12.16}}.

[0096] The serialized data exchange objects are published and synchronized through standardized application programming interfaces in an event-driven or request-response pattern.

[0097] Among them, the event-driven model refers to the system actively pushing messages to subscribed consumers when new data is generated; the request-response model refers to the interaction method in which consumers actively initiate requests and the system returns corresponding data.

[0098] Specifically, the system can deploy a message broker (such as RabbitMQ or Apache Kafka) to implement event-driven publishing, publishing new data exchange objects as messages to specific topics or exchanges. Simultaneously, the system also provides a RESTful API endpoint supporting a request-response model. When an HTTP GET request is received, it returns the latest or a data exchange object for a specified time period. For example, in event-driven mode, whenever a new multi-dimensional collaborative metric is calculated, the system immediately publishes its serialized JSON string to a RabbitMQ exchange named `building.energy.metrics`. All applications subscribed to this exchange (such as scheduling systems) will receive this message in real time. In request-response mode, the building energy efficiency management platform can send an HTTP GET request to https: / / XXXXXX / XX / XXXXXX at any time and receive the same JSON data in the response body.

[0099] In the photovoltaic-storage-direct-flexible collaborative scheduling optimization model, data exchange objects are subscribed to and parsed through interface protocol specifications to drive the generation of power control commands.

[0100] Among them, the photovoltaic-storage-direct-flexible coordinated scheduling optimization model refers to a mathematical optimization model or decision algorithm that calculates the optimal operating power of photovoltaic, energy storage, and flexible load equipment with the objectives of economy, low carbon emissions, or reliability.

[0101] Specifically, this optimization model, acting as a client, subscribes to relevant topics via a message queue client or periodically calls a RESTful API according to the interface protocol specifications. Upon receiving a data exchange object, the client first parses the JSON, extracting key indicators such as net_power_kW (net power) and flex_value_rmb_per_h (regulation value). The optimization model uses these real-time indicators as one of its input parameters, combined with constraints such as electricity price and equipment status, to solve an optimization problem, ultimately generating a sequence of power setpoint instructions for photovoltaic inverters and energy storage converters. For example, after subscribing to a message, the scheduling optimization model parses the current net_power_kW = ... kW (power supplied to the grid), flex_value_rmb_per_h= The current electricity price is ¥ / h (revenue status), but a surge in electricity prices is predicted for the next period. The model, through calculations, decides to increase the energy storage charging power, adjusting net_power_kW to [value missing]. kW, and generate the instruction: {"pv_curtail":0%,"ess_charge_power_kw":10,"grid_feedin_kw":5} and send it to the device for execution.

[0102] In the building lifecycle carbon efficiency management model, data exchange objects are subscribed to and parsed through interface protocol specifications to trigger carbon flow accounting and energy efficiency assessment processes.

[0103] Among them, the building life cycle carbon efficiency management model refers to a system or module that calculates, analyzes, evaluates and predicts energy consumption and carbon emissions during the building operation phase, and formulates carbon reduction strategies.

[0104] Specifically, the management model periodically (e.g., hourly) calls the API to obtain the latest indicator data according to the interface protocol specifications. The model parses the data exchange object, extracting indicators such as `carbon_intensity_kgco2_per_h` (carbon emission intensity) and `system_efficiency` (system efficiency). The model adds the current carbon emission intensity to the daily, monthly, and yearly cumulative carbon emissions and updates the building carbon efficiency dashboard by combining this with static carbon data from the Building Information Modeling (BIM). Simultaneously, the model compares the current system efficiency with historical benchmarks or benchmark values ​​for similar buildings to perform energy efficiency assessments. If a decline in efficiency is detected, an energy efficiency diagnostic analysis is triggered. For example, the carbon efficiency management model calls the API at 14:30 and parses the carbon emission intensity for the past hour to obtain... The model includes this value in "Today's Cumulative Carbon Reduction," updating the figure accordingly. Meanwhile, the model detected that the system efficiency was 91%, which was lower than the set baseline of 94%. Therefore, it automatically generated an energy efficiency alarm and initiated correlation analysis to check whether it was related to the decrease in efficiency of specific equipment (such as chillers).

[0105] In the source-grid-load-storage cluster collaborative control model, data exchange objects are subscribed to and parsed through interface protocol specifications.

[0106] Among them, the source-grid-load-storage cluster collaborative control model refers to the centralized optimization and control model used by the regional power grid dispatch center to coordinate and manage multiple distributed "source-grid-load-storage" resources within its jurisdiction, in order to support the safe, economical, and low-carbon operation of the power grid.

[0107] Specifically, this control model, as a regional application, aggregates indicator data from multiple buildings within its jurisdiction. The model subscribes to the `building.energy.metrics` topic published by each building or makes batch calls to the APIs of each building through an interface protocol specification. The model parses the data exchange objects reported by each building, extracting key indicators such as `net_power_kW` (which can be considered as adjustable net load) and `flex_value_rmb_per_h` (adjustment value). Based on the net power and adjustment potential of all buildings, the model performs cluster-level optimization calculations to determine whether to issue demand response commands to certain buildings or adjust the grid operation mode. For example, the regional dispatch center subscribes to the indicators of three buildings, A, B, and C. The parsed data shows: Building A `net_power_kW` = -15, Building B `net_power_kW` = 20, and Building C `net_power_kW` = 5. The cluster model calculates that the regional net load is 10kW, but the grid needs to reduce the load by 20kW. Based on the adjustment value of each building, the model issues load reduction instructions to building B (a major electricity user) and building C, while building A (which is generating electricity) remains unchanged.

[0108] Therefore, according to the above implementation method, the system can provide the fusion indicators obtained by internal analysis and calculation to various upper-level optimization and control models in a standard, open and interoperable manner, realizing efficient and reliable decoupling and collaboration between the monitoring system and decision-making applications, and enabling the value of panoramic monitoring data to be fully released in applications at different levels.

[0109] In some embodiments, the above method further includes a continuous optimization step for key models and parameters in the data fusion computing pipeline, the continuous optimization step including: A supervised training sample set is constructed based on historical time series data. Supervised training is performed on the multimodal deep learning model in the deep fusion inference engine, and the convergence of the validation set loss function is used as the trigger condition for terminating the model training.

[0110] The supervised training sample set refers to the set of data pairs compiled from historical monitoring data, which contains input features and corresponding true labels (or target values), and is used to train the model to learn the mapping relationship from input to output. The convergence of the validation set loss function means that when a portion of the reserved samples (validation set) are input into the model being trained, the calculated prediction error (loss) value no longer decreases significantly in multiple consecutive training iterations, indicating that the model may have learned the main patterns in the data, and continued training may lead to overfitting of the training data.

[0111] Specifically, the system can extract historical data from the past 12 months from the spatiotemporally aligned canonical dataset. The data from the first 11 months is used as the training set, and the data from the last month is split into a validation set and a test set in chronological order. The input feature (X) for each sample can be multi-dimensional time-series data from the past hour (such as photovoltaic power, load, and temperature) and static model parameters, while the label (y) can be the system's comprehensive energy efficiency value or the actual measured value of carbon emission intensity at the next time step. The system uses Mean Squared Error (MSE) as the loss function and employs a stochastic gradient descent optimizer for model training. During training, the MSE loss on the validation set is continuously monitored. When the loss does not decrease for 20 consecutive training epochs, an early stopping mechanism is triggered, terminating training and saving the current model parameters. For example, the system uses data from the entire year of 2022 to construct the sample set. The dimensions of the input feature X are [batch size, 60, 10], indicating that each sample contains a historical sequence with 60 time steps (1 hour) and 10 feature dimensions. The label y is a scalar, such as 0.85 (representing energy efficiency). After 100 epochs of model training, the validation set loss decreased from the initial 0.05 to 0.004 and stabilized. At epoch 85, the system detected that the validation set loss had not decreased for 20 consecutive epochs (epochs 65 to 84), so it stopped training at epoch 85 and used the model parameters obtained at epoch 64 (corresponding to the minimum point of validation loss) as the final model.

[0112] An online learning and model fine-tuning pipeline for multimodal deep learning models is constructed. Online monitoring data streams are collected periodically, and the model's weight parameters are updated using a small-batch incremental training method to adapt to changes in building operation status and equipment configuration.

[0113] Among them, the online learning and model fine-tuning pipeline refers to an automated data processing and model retraining process that can periodically add newly generated data during system operation to the training process and make minor adjustments to the existing model; the small-batch incremental training method refers to using only a small portion of the latest data to fine-tune the model weights each time while retaining the original knowledge of the model, so as to quickly adapt to new trends and avoid forgetting old patterns.

[0114] Specifically, the system can be configured with a scheduled task (e.g., executed quarterly). Upon startup, this task first collects new monitoring data generated in the past quarter. This data undergoes the same preprocessing steps as when building the historical sample set, resulting in a new mini-batch of training samples. Then, the system loads the deployed multimodal deep learning model, freezing most of the parameters of its bottom-level feature extraction layers and unfreezing only the parameters of a few fully connected layers near the output layer. Using the new mini-batch data, the model is trained for a few epochs at a low learning rate, allowing for fine-tuning of the model's decision rules. After validation, the updated model replaces the old online model via hot deployment. For example, the system launched its online learning pipeline at the beginning of Q4 2023. It collected new data from July to September 2023, generating approximately 2000 new samples. The system loaded the current online model, freezing all parameters, including the Transformer encoder and GCN graph convolutional layers, and fine-tuning only the last two fully connected layers. Using this new data, it trained for 5 epochs at an initial learning rate of 0.0001. Validation shows that the fine-tuned model is more accurate in predicting the localized load shift patterns that emerged in the third quarter due to building renovations, while its predictive performance for historical overall patterns remains unchanged.

[0115] Deploy an external data integration pipeline to dynamically maintain the carbon emission factor knowledge base and the time-of-use electricity pricing strategy knowledge base for the collaborative fusion computing module, and perform real-time mapping and updates to the knowledge base through the pipeline.

[0116] Among them, the external data integration pipeline refers to a dedicated data process used to automatically acquire, parse, clean and import data from third-party data sources outside the system; the carbon emission factor knowledge base and the time-of-use electricity pricing strategy knowledge base refer to a structured database or configuration file that stores real-time or predicted carbon emission factors of different regional power grids, as well as electricity prices for different time periods and different user types.

[0117] Specifically, the system can be configured with multiple data collectors to periodically access external application programming interfaces (APIs) such as the carbon emission factor platform released by the provincial ecological and environmental departments and the power trading center information platform released by the power grid company. The pipeline calls these APIs according to preset rules (e.g., every 15 minutes) to obtain the latest data. The acquired raw data, after format parsing and validity verification, is converted into an internally unified format and updated to the corresponding knowledge base database tables. During calculation, the collaborative fusion calculation module directly queries the required factors and prices from the latest knowledge base. For example, if the external data integration pipeline calls the East China Power Grid carbon emission factor API at 14:15 and receives the response {“region”:“east_china”,“timestamp”:“2023-10-27T14:15:00Z”,“carbon_factor”:0.521}, it indicates that the current carbon emission factor is 0.521 kg of carbon dioxide per kilowatt-hour (…). Simultaneously, the pipeline calls the electricity trading API to obtain the time-of-use pricing strategy for the day: {"peak":["10:00-12:00","14:00-16:00","19:00-21:00"],"valley":["00:00-08:00"],"peak_price":1.2,"valley_price":0.4}, indicating a peak-hour price of ¥1.2 per kWh and a valley-hour price of ¥0.4 per kWh. The pipeline updates this data to the knowledge base. During the collaborative calculation at 14:30, the module queries the knowledge base and finds that 14:30 is a peak time, and the price of ¥1.2 per kWh should be used for calculation.

[0118] Therefore, according to the above implementation method, the system enables the data fusion computing pipeline to have self-evolution and dynamic adaptation capabilities. By periodically retraining the model with new data, it captures the changing trends of the operating status and synchronizes key external parameters in real time, thereby ensuring that the accuracy of panoramic monitoring and the effectiveness of decision support are maintained continuously during the long-term operation of the system.

[0119] In some embodiments, the above method further includes supporting steps for ensuring the credibility of the entire data flow, the real-time performance of processing, and the resilience of the system. These supporting steps include: Construct an edge-cloud collaborative computing plane, instantiate shallow fusion services on edge computing nodes, and deploy deep fusion inference engines and collaborative fusion computing modules in cloud data centers to form a low-latency data processing pipeline and a high-computing-power model inference pipeline.

[0120] The edge-cloud collaborative computing plane refers to a logically unified and physically distributed computing resource scheduling and management layer. It abstracts the computing, storage and network resources on the edge side and the cloud side, enabling applications to perform different computing tasks in the most suitable location, just like using a single computing platform.

[0121] Specifically, the system can manage this compute plane by defining different namespaces and node selectors through a container orchestration platform (such as Kubernetes). Node selectors are configured for the "edge" namespace to ensure that latency-sensitive Pods (container groups) requiring local data processing, such as "shallow fusion services," are scheduled to edge servers tagged with "edge=true". Selectors are configured for the "cloud" namespace to ensure that services requiring significant GPU (Graphics Processing Unit) computing power, such as "deep fusion inference engines," are scheduled to GPU servers in the cloud data center. For example, in a deployment in an industrial park, the system deploys one edge server (model: IEITANK-870, CPU: Intel i7, tag: edge=true) in each of three power distribution rooms. A cluster of four NVIDIA A100 GPU servers is located in the cloud data center. Kubernetes schedules three instances of the shallow fusion service to the three edge servers respectively, processing second-level data from the devices in the power distribution rooms in real time. An instance of the deep fusion inference engine is scheduled to a cloud GPU server to periodically receive fused data from all edges and perform batch deep inference every 15 minutes.

[0122] A multi-layered data trust assurance chain is constructed, which includes real-time consistency verification, cross-source logical verification, and periodic manual review. Real-time deviation detection and correlation verification are performed on monitoring data streams that are from the same source or have physical constraints. Based on the residual sequence of the verification output, the hyperparameters of the fusion algorithm in the data fusion computing pipeline are dynamically adjusted.

[0123] Among them, real-time consistency verification refers to the process of comparing instantaneous values ​​of data from different sensors of the same device or multiple data sources that should be physically consistent at the moment of data acquisition or processing; cross-source logic verification refers to the process of verifying the correlation of data by using physical laws or business logic relationships (such as energy conservation) between different data sources; periodic manual review refers to the process of manually comparing key indicators output by the system with offline records, instrument readings or other reliable data by maintenance personnel on a regular basis.

[0124] Specifically, the system can embed real-time consistency verification logic within the edge data acquisition agent. For example, using the principle of energy conservation, it can perform real-time verification of photovoltaic power generation, energy storage charging and discharging power, building load power, and power exchanged with the grid simultaneously. If their algebraic sum (considering fixed losses) continuously deviates from the theoretical threshold, a data reliability alarm is triggered, and the fusion algorithm is notified to adjust the weights of the corresponding data sources. In the cloud, a cross-source logic verification service can be deployed, for example, based on the formula... (Grid power + Photovoltaic power = Load power + Loss power) Verify the real-time power data and calculate the residual. If the standard deviation of the residual sequence remains consistently high, it indicates a potential systematic error in a data source. The verification service can issue an alarm and trigger an adaptive Kalman filter algorithm to automatically increase the observation noise covariance R value for the suspected data source, reducing its weight in the fusion process. Monthly, maintenance personnel export the total monthly power generation from the system and manually verify it against the values ​​displayed on the photovoltaic inverter's built-in statistics screen, recording any discrepancies. For example, if the system detects an error during the 10:00-10:15 period, calculated based on meter and inverter data... The mean ratio is calculated based on load monitoring. The mean is 8 kW higher, and the standard deviation of the residual sequence is relatively large. The cross-source logic verification service determines that there is a logic bias, triggers a diagnosis of the load monitoring sensor network, and notifies the adaptive filtering algorithm to temporarily increase the estimated value R of the load power data observation from 0.1 to 0.3 within the next hour, thereby reducing its confidence in the fusion calculation until the bias is eliminated.

[0125] Enable role-based access control engine and data masking filter to perform business role-based aggregation calculation and sensitive information masking on fine-grained energy consumption and carbon emission data streams output through standardized application programming interfaces.

[0126] Among them, the role-based access control engine is a permission management model that assigns system access permissions to different roles and then grants roles to users, rather than directly granting permissions to users; the data desensitization filter is a data processing component that transforms the data flowing through it in real time to hide or obscure sensitive information.

[0127] Specifically, the system can integrate a lightweight access control list service. Roles such as "System Administrator," "Energy Analyst," "Tenant," and "Visitor" are defined, and each role is configured with accessible application programming interface (API) endpoints and data fields. When an external request arrives at the API gateway, the gateway verifies the role associated with the requester's token. Simultaneously, before data serialization and output, it passes through a data masking filter. This filter determines the granularity of the output data based on the requester's role: for the "Tenant" role, only aggregated energy consumption for their leased area is returned; for the "Energy Analyst" role, fine-grained data at the building level but not the device level is returned, and commercially sensitive peak load times are obfuscated (e.g., "14:28" is masked as "14:20~14:30"). For example, user A (role: Energy Analyst) requests the building's hourly energy consumption for yesterday. The access control engine allows them access to " / api / v1 / building / hourly-energy". Before returning the data, the data masking filter aggregates the detailed energy consumption of each room into floor-level energy consumption and masks the specific device identifiers. User B (role: visitor) requests the same interface, but is rejected by the access control engine and instead can only access a read-only interface " / api / v1 / building / today-total" that returns the building's total energy consumption for the day.

[0128] On the communication plane consisting of edge data acquisition agents, edge computing nodes, and cloud data centers, deploy network intrusion detection systems and secure encrypted transport layer channels.

[0129] Among them, a network intrusion detection system is a software and hardware system that monitors network traffic and attempts to identify and warn of potential malicious activities or policy violations; a transport layer secure encrypted channel refers to a virtual communication path established between two communication nodes that uses protocols such as TLS / SSL to encrypt transmitted data in order to prevent eavesdropping and tampering.

[0130] Specifically, the system can deploy a hardware or software intrusion detection system at the cloud data center network ingress to analyze all network packets from the edge in real time and match them with known attack signatures. Simultaneously, transport layer security protocols are enforced on all network communication links. TLS 1.3 encrypted channels are established between the edge data acquisition agent and edge computing nodes, and between edge nodes and the cloud, using certificate-based two-way authentication. All data transmitted via message queues (such as MQTT) must also be transmitted over the TLS channel. For example, if the edge data acquisition agent (IP: 192.168.1.50) needs to send data to the cloud message agent (domain: mqtt.cloud.com), first, the agent and the cloud server complete a TLS handshake, and both parties verify their certificates. Subsequently, all MQTT protocol messages are transmitted in this encrypted tunnel. The cloud-deployed network intrusion detection system, at the decrypted application layer, detects an MQTT message from an edge agent containing an abnormally long subject name, matching the characteristics of a "buffer overflow attack detection," and immediately generates an alarm and temporarily blocks the connection to that agent.

[0131] Configure a data lake backup strategy and disaster recovery process with version control, perform periodic snapshot backups on standardized data sets and multi-dimensional collaborative fusion indicators, and preset system-level recovery time targets and data recovery point targets.

[0132] Among them, the version-controlled data lake backup strategy refers to backing up not only the current data when backing up the data lake (which stores raw and standardized data), but also retaining copies of the data's state at multiple historical points in time for retrospective or recovery; the system-level recovery time target refers to the maximum time allowed from system crash to the restoration and availability of core business functions after a disaster; the data recovery point target refers to the amount of data loss that the system can tolerate during recovery, i.e., the maximum point in time before the disaster that the data can be restored to.

[0133] Specifically, the system can be configured with backup management software to implement differentiated backup strategies for the database storing "temporally aligned canonical datasets" and the time-series database storing "multidimensional collaborative fusion indicators." For the canonical datasets, a full snapshot backup is performed daily at midnight, retaining the most recent 30 days' versions. For the multidimensional collaborative fusion indicators, incremental backups are performed hourly. Full backup data is synchronized to an off-site backup storage center. The recovery process documentation clearly stipulates that in the event of a complete failure of the primary cloud data center, core data query and latest indicator calculation services should be restored within 30 minutes in the backup data center, ensuring that the restored data loses at most the most recent 5 minutes of new data. For example, the system performs a full snapshot of the Hive tables of the canonical datasets at 02:00 daily, with the snapshot version marked as 20231027. The off-site backup center stores 30 full snapshots from 20230928 to 20231027. Simultaneously, the multidimensional collaborative fusion indicator database (such as InfluxDB) will be updated with new data backups every hour. At 10:15 AM one day, the primary data center went down. Disaster recovery was initiated. In the backup data center, using a full snapshot of version 20231027 and incremental backups of metrics up to 10:10 AM, the database service was restored within 25 minutes. After system recovery, the data was complete, but the latest data time for the multi-dimensional collaborative fusion metrics was 10:10 AM, achieving the design goals of a 30-minute recovery time and a 5-minute data recovery point.

[0134] Therefore, based on the above implementation method, the system can build a complete and in-depth technical assurance system from five dimensions: computing architecture, data quality, access security, network security and business continuity, to ensure that the panoramic monitoring method can operate continuously, reliably and securely in complex and open engineering environments, and provide a solid and reliable data foundation for business decisions.

[0135] Figure 3 This is a structural block diagram of a panoramic monitoring system based on optical storage, direct current and flexible control, according to an embodiment of the present invention.

[0136] like Figure 3 As shown, this panoramic monitoring system based on optical storage, direct current, and flexible operation includes: The data acquisition layer construction module 210 is used to configure multi-dimensional heterogeneous data sources covering photovoltaic storage direct current and flexible equipment, building body, environment and power grid side, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer.

[0137] The data set generation module 220 is used to sequentially perform data cleaning, scale normalization, timestamp synchronization and spatial registration operations on the monitoring data collected by the distributed panoramic perception data acquisition layer, and generate a standardized data set that is spatiotemporally aligned.

[0138] The computational pipeline construction module 230 is used to build and execute a collaborative data fusion computational pipeline from the edge to the cloud. The pipeline includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge, a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud, and a collaborative fusion stage that performs coupled computation on system operating status and energy and carbon indicators.

[0139] The panoramic monitoring result generation module 240 is used to publish the fusion features and indicators output in the collaborative fusion stage through a standardized application programming interface.

[0140] The specific functions and examples of each module and submodule of the device in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0141] According to embodiments of the present invention, the above-described method of the present invention can be applied to a computer device and a readable storage medium.

[0142] Figure 4 A schematic block diagram of a computer device 600 that can be used to implement embodiments of the present invention is shown. The computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0143] like Figure 4 As shown, the computer device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the computer device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0144] Multiple components in computer device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows computer device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0145] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as a panoramic monitoring method based on optical storage, direct current, and flexible positioning. For example, in some embodiments, a panoramic monitoring method based on optical storage, direct current, and flexible positioning can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the computer device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of a panoramic monitoring method based on optical storage, direct current, and flexible positioning described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform a panoramic monitoring method based on optical storage, straightness, and flexibility.

[0146] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0147] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0148] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0149] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0150] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0151] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0152] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0153] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A panoramic monitoring method based on light storage direct flexible, characterized in that, include: Configure multi-dimensional heterogeneous data sources covering photovoltaic-storage-direct-flexible equipment, building structure, environment and power grid, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer; The monitoring data collected by the distributed panoramic perception data acquisition layer is sequentially subjected to data cleaning, scale normalization, timestamp synchronization and spatial registration operations to generate a standardized data set that is aligned in time and space. Construct and execute a collaborative data fusion computing pipeline from edge to cloud, which includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge, a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud, and a collaborative fusion stage that performs coupled computation on system operating status and energy and carbon indicators. The fusion features and indicators output from the collaborative fusion phase are published through a standardized application programming interface.

2. The method of claim 1, wherein, The configuration covers multi-dimensional heterogeneous data sources including photovoltaic-storage-direct-drive-flexible equipment, building structures, the environment, and the power grid, and deploys edge data acquisition agents to construct a distributed panoramic perception data acquisition layer, including: Configure heterogeneous data source access endpoints covering the operating status of photovoltaic-storage-direct-flexible equipment, building IoT sensor nodes, building information models, power grid interaction interfaces, and third-party environmental data platforms; The edge data acquisition agent is deployed at the physical proximity of the optical storage direct-flexible device and the building Internet of Things sensor network. The edge data acquisition agent is used to perform data extraction, transformation and loading processes on the raw data stream originating from the local area, and temporarily store it in the edge data lake. Define the differentiated data transmission service quality level and routing strategy for the communication link between the edge data acquisition agent and the cloud data center; The metadata registry of the heterogeneous data source access endpoint and the edge data acquisition agent is maintained in the cloud data center.

3. The method of claim 1, wherein, The monitoring data collected by the distributed panoramic perception data acquisition layer is sequentially subjected to data cleaning, scale normalization, timestamp synchronization, and spatial registration operations to generate a standardized data set aligned in time and space, including: Instantiate the data cleaning computation graph and sequentially perform anomaly detection and filtering based on statistical distribution and isolated forest, as well as missing value imputation based on interpolation algorithm on the input monitoring data stream; Perform feature scaling or standardization transformations on the cleaned data sequence; Invoke the spatiotemporal alignment service to perform upsampling or downsampling operations on heterogeneous time-series data streams at a preset reference sampling frequency, and perform coordinate system one and spatial index construction on static or quasi-static model data with spatial attributes; The data units that have undergone feature scaling and spatiotemporal alignment are partitioned and cataloged according to the data pattern of the static or quasi-static model data and their respective business domains, and persistently stored as the spatiotemporally aligned standardized data set.

4. The method of claim 1, wherein, The steps of building and executing a collaborative data fusion computing pipeline from edge to cloud include: A shallow fusion service is instantiated at the edge computing node. The shallow fusion service is used to perform adaptive filtering and fusion calculations on the real-time high-dimensional data stream from the distributed panoramic perception data acquisition layer and output low-dimensional fusion temporal features. A deep fusion inference engine is deployed in a cloud data center. The deep fusion inference engine receives the low-dimensional fusion temporal features and static or quasi-static model parameter data from the standard dataset. It performs feature extraction and association learning on heterogeneous data streams through a multimodal deep learning model to generate system-level deep fusion feature vectors. A collaborative fusion computing module is deployed in the cloud data center. Based on the deep fusion feature vector, the collaborative fusion computing module performs coupled calculations of operating status, carbon emission intensity and flexible adjustment value to generate multi-dimensional collaborative fusion indicators. The multidimensional collaborative fusion indicators are serialized into a standardized data exchange format and published through the standardized application programming interface.

5. The method of claim 1, wherein, The step of publishing the fusion features and indicators output from the collaborative fusion phase through a standardized application programming interface includes: Define data models and interface protocol specifications for the multi-dimensional collaborative fusion indicators output in the collaborative fusion phase; The multi-dimensional collaborative fusion indicators are serialized into structured data exchange objects according to the data pattern; The serialized data exchange object is published and synchronized through the standardized application programming interface in an event-driven or request-response mode. In the optical-storage-direct-flexible collaborative scheduling optimization model, the data exchange object is subscribed to and parsed through the interface protocol specification to drive the generation of power control commands; In the building life cycle carbon efficiency management model, the data exchange object is subscribed to and parsed through the interface protocol specification to trigger the carbon flow accounting and energy efficiency assessment process; In the source-grid-load-storage cluster collaborative control model, the data exchange object is subscribed to and parsed through the interface protocol specification.

6. The method of claim 1, wherein, The method further includes a continuous optimization step for key models and parameters in the data fusion computing pipeline, the continuous optimization step including: A supervised training sample set is constructed based on historical time series data. Supervised training is performed on the multimodal deep learning model in the deep fusion inference engine, and the convergence of the validation set loss function is used as the trigger condition for terminating the model training. An online learning and model fine-tuning pipeline for the multimodal deep learning model is constructed, online monitoring data streams are collected periodically, and the weight parameters of the multimodal deep learning model are updated in a small-batch incremental training manner. An external data integration pipeline is deployed to dynamically maintain the carbon emission factor knowledge base and the time-of-use electricity pricing strategy knowledge base for the collaborative fusion computing module, and to perform real-time mapping and updates on the carbon emission factor knowledge base and the time-of-use electricity pricing strategy knowledge base through the external data integration pipeline.

7. The method of claim 1, wherein, The method also includes supporting steps to ensure the credibility of the entire data process, the real-time performance of processing, and the resilience of the system. These supporting steps include: Construct an edge-cloud collaborative computing plane, instantiate shallow fusion services on edge computing nodes, and deploy deep fusion inference engines and collaborative fusion computing modules in cloud data centers to build a low-latency data processing pipeline and a high-computing-power model inference pipeline; A multi-layered data trust assurance chain is constructed, which includes real-time consistency verification, cross-source logical verification, and periodic manual review. Real-time deviation detection and correlation verification are performed on monitoring data streams that are from the same source or have physical constraints. Based on the residual sequence of the verification output, the hyperparameters of the fusion algorithm in the data fusion calculation pipeline are dynamically adjusted. Enable role-based access control engine and data desensitization filter to perform business role-based aggregation calculation and sensitive information masking on the fine-grained energy consumption and carbon emission data stream output through the standardized application programming interface; On the communication plane formed by the edge data acquisition agent, the edge computing node, and the cloud data center, a network intrusion detection system and a transport layer secure encryption channel are deployed; Configure a data lake backup strategy and disaster recovery process with version control, perform periodic snapshot backups on the standardized data set and multi-dimensional collaborative fusion indicators, and preset system-level recovery time targets and data recovery point targets.

8. A panoramic monitoring system based on light storage direct flexible, characterized in that, include: The data acquisition layer construction module is used to configure multi-dimensional heterogeneous data sources covering photovoltaic storage direct current and flexible equipment, building body, environment and power grid side, and deploy edge data acquisition agents to build a distributed panoramic perception data acquisition layer. The data set generation module is used to sequentially perform data cleaning, scale normalization, timestamp synchronization and spatial registration operations on the monitoring data collected by the distributed panoramic perception data acquisition layer to generate a standardized data set that is aligned in time and space. The computational pipeline construction module is used to build and execute a collaborative data fusion computational pipeline from the edge to the cloud. The pipeline includes: a shallow fusion stage that performs adaptive filtering on real-time data streams at the edge, a deep fusion stage that performs deep learning feature extraction on cross-modal data in the cloud, and a collaborative fusion stage that performs coupled computation on system operating status and energy and carbon indicators. The panoramic monitoring result generation module is used to publish the fusion features and indicators output by the collaborative fusion stage through a standardized application programming interface.

9. A computer device, comprising: include: At least one processor; and a memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein the computer instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1-9. in, Computer instructions are used to cause a computer to perform the method according to any one of claims 1-7.

Images