An AI-based digital energy carbon management system

Abstract

The application discloses an AI-based digital energy-carbon management system, and belongs to the technical field of comprehensive energy and carbon emission collaborative control. The system adopts a four-level architecture of "perception-edge-platform-application", and initiates a core technical route of "energy-carbon dual-domain feature engine+knowledge graph event driving+meta-learning fusion". The system is internally provided with four core functional modules of AI load prediction, carbon emission inversion, self-explaining abnormality tracing and MILP low-carbon scheduling, and can support the access of multiple energy media such as electricity, cold, heat, gas and hydrogen. The system supports multi-scene "zero code" rapid replication, has a short deployment cycle, and can reduce comprehensive carbon emission and improve annual economic benefits after application. The application solves the technical pain points of the existing energy-carbon management system, such as data fragmentation, poor timeliness, low prediction accuracy and difficult scheduling closed loop, has outstanding novelty, creativity and practicability, and can be widely applied to various comprehensive energy scenes such as parks, factories and public buildings.

Description

Technical Field

[0001] This invention relates to the field of integrated energy and carbon emission synergistic control technology, and in particular to an AI-based digital energy and carbon management system. Background Technology

[0002] Against the backdrop of advancing the "dual carbon" goals, various industrial parks, factories, and public buildings have an increasingly urgent need for efficient energy utilization and precise carbon emission control, making energy-carbon management systems a core vehicle for achieving these goals. However, existing energy-carbon management systems generally suffer from the following technical challenges:

[0003] 1. Severe data fragmentation: Most systems only manage a single energy medium (such as electricity), and cannot achieve collaborative data collection and fusion analysis of multiple media such as electricity, cold, heat, gas and hydrogen. In addition, carbon emission data is disconnected from energy data, making it difficult to form "energy-carbon synergistic" management and control.

[0004] 2. Insufficient timeliness and accuracy: The traditional mode of "post-event accounting and manual reporting" results in delayed data updates, low load forecast accuracy, and large carbon emission accounting deviations, which cannot meet the needs of minute-level carbon emission control and real-time scheduling.

[0005] 3. Difficulty in tracing the source of anomalies: When carbon emissions are abnormal, it is impossible to quickly locate the root cause of the anomaly, and there is a lack of interpretability, which brings great inconvenience to the operation and maintenance personnel in troubleshooting the problem;

[0006] 4. Lack of closed-loop scheduling: The lack of a scientific low-carbon scheduling model makes it impossible to achieve an end-to-end closed loop of "prediction-accounting-optimization-scheduling", resulting in limited energy utilization efficiency and carbon emission reduction effects.

[0007] 5. Poor replicability: System deployment in different scenarios requires a lot of customized coding, resulting in long deployment cycles, high costs, and difficulty in rapid promotion and application.

[0008] Therefore, this invention proposes a digital and intelligent energy and carbon management system that features multi-media collaboration, high-precision prediction, rapid anomaly tracing, closed-loop low-carbon scheduling, and high replicability. Summary of the Invention

[0009] To address the technical pain points of existing energy and carbon management systems, such as data fragmentation, poor timeliness, low prediction accuracy, difficulty in scheduling closed loops, and poor replicability, this invention provides an AI-based intelligent energy and carbon management system that enables multi-energy collaboration, energy and carbon integration, precise control, and rapid replication, thereby improving energy utilization efficiency, reducing carbon emissions, and increasing economic benefits.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0011] An AI-based intelligent energy and carbon management system adopts a four-level architecture of "sensing-edge-platform-application", including:

[0012] a) Sensing layer, used for synchronous acquisition of multi-energy flow data, covering at least five energy media: electricity, cooling, heating, gas and hydrogen, with a sampling period of ≤1min, and simultaneously acquiring data related to meteorology, electricity price, production plan, maintenance order and dynamic carbon emission factor;

[0013] b) Edge layer, with built-in edge computing gateway, to complete the functions of cleaning, protocol conversion, local caching and breakpoint resume of collected data, where the local caching time is ≥48h;

[0014] c) At the platform layer, a microservice containerized deployment approach is adopted, including at least:

[0015] The AI ​​load forecasting module is used to output load forecast curves at multiple time scales, including day-ahead, intraday, and real-time forecasts. The mean absolute percentage error (MAPE) of the day-ahead load forecast is ≤3.6%.

[0016] The carbon emission inversion module is used to calculate the carbon emission curve based on the predicted load output by the AI ​​load prediction module and the dynamic carbon emission factor collected by the sensing layer. Its carbon emission calculation deviation is ≤±3.3%.

[0017] The self-explaining anomaly tracing module is used to locate the top-3 high-contribution anomalies when the deviation between the actual and calculated carbon emission values ​​exceeds a preset threshold, with an anomaly tracing time ≤ 1 minute.

[0018] The MILP low-carbon scheduling module aims to minimize the cost of electricity purchase, carbon emission cost, and energy storage lifetime penalty. It generates low-carbon scheduling strategies using a 15-minute rolling solution method, with a solution time of ≤30 seconds.

[0019] d) Application layer, used to provide a visual cockpit, automatic generation of carbon emission reports and third-party API interfaces, with a configuration time of ≤2 days for new scenario access.

[0020] Furthermore, the AI ​​load prediction module adopts a parallel base model of "XGBoost+TCN+GRU", and fuses the knowledge graph event confidence vector Evt and the equivalent carbon load through the LightGBM meta-learner. Where the dimension k of the Evt vector is ≤32, the formula for calculating the equivalent carbon load LCE(t) is:

[0021]

[0022] In the formula For the real-time load of the i-th energy medium, is the dynamic carbon emission factor for the i-th energy medium.

[0023] Furthermore, the carbon emission inversion module uses the following calculation formula to calculate carbon emissions:

[0024]

[0025] In the formula Let be the total carbon emissions at time t. Let be the energy conversion efficiency of the i-th energy medium. This is an event correction item used to correct the impact of various unforeseen events on carbon emission accounting results.

[0026] Furthermore, the MILP low-carbon scheduling module linearizes the energy storage cycle life loss, achieving a linear fit goodness of ≥0.95, and supports a two-charge, two-discharge scheduling strategy for energy storage with a scheduling time granularity of ≤15min.

[0027] Furthermore, the platform layer supports horizontal scaling, with a single-node fault recovery time (RTO) of ≤5 min, meeting the requirement for continuous and stable operation 24 / 7.

[0028] An AI-based intelligent energy and carbon management method, utilizing the aforementioned AI-based intelligent energy and carbon management system, sequentially completes the following steps:

[0029] S1: The sensing layer completes the synchronous acquisition of multi-energy medium load and related data, and uploads it to the edge layer;

[0030] S2: The edge layer cleans and converts the collected data according to the protocol, caches it locally and uploads it synchronously to the platform layer. Local caching is completed when the network is disconnected, and interrupted transmission is resumed after the network is restored.

[0031] S3: The platform layer completes multi-timescale load forecasting through the AI ​​load forecasting module, completes carbon emission accounting through the carbon emission inversion module, completes carbon emission anomaly location through the self-explaining anomaly tracing module, and generates rolling low-carbon scheduling strategies through the MILP low-carbon scheduling module.

[0032] S4: The application layer receives data and strategies output from the platform layer, displays them in a visual cockpit, generates carbon emission reports, or synchronizes them to other systems via third-party API interfaces.

[0033] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI-based intelligent energy and carbon management method of claim 6.

[0034] The beneficial effects of this invention are as follows:

[0035] 1. Significantly improved accuracy in energy and carbon coordinated management: The pioneering technical approach of "energy-carbon dual-domain feature engine + knowledge graph event-driven + meta-learning fusion" enables collaborative acquisition of multi-energy media and deep fusion of energy and carbon data, achieving a day-ahead load forecast MAPE ≤ 3.6% and carbon emission accounting deviation ≤ ±3.3%;

[0036] 2. Real-time closed-loop scheduling with outstanding carbon emission reduction effect: Constructing an end-to-end closed loop of "prediction-accounting-source tracing-scheduling" to achieve minute-level anomaly source tracing and 15-minute rolling low-carbon scheduling. After application, the overall carbon emissions are reduced, meeting the "dual carbon" target management requirements.

[0037] 3. Zero-code rapid replication and high deployment efficiency: Supports zero-code configuration for multiple scenarios, with a deployment cycle of ≤2 days, solving the pain points of existing systems with high degree of customization and poor replicability, and can be quickly promoted to various parks, factories and public buildings;

[0038] 4. Significant economic benefits and strong practicality: By optimizing electricity purchase costs and extending energy storage life through low-carbon dispatch, the annual economic benefits are greatly improved. At the same time, the system is containerized and supports horizontal expansion. The single node fault recovery time is ≤5min, which meets the requirements of 7×24h continuous operation and is suitable for various industrial and commercial scenarios.

[0039] 5. Strong self-explanatory power and reduced operation and maintenance costs: The SHAP value anomaly tracing algorithm can quickly locate the root cause of anomalies and provide interpretable evidence, reducing the troubleshooting time of operation and maintenance personnel and reducing operation and maintenance costs. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a diagram illustrating the four-level architecture and data flow of the system of this invention.

[0042] Figure 2 Internal structure of the AI ​​load forecasting module;

[0043] Figure 3 Flowchart for carbon emission inversion and anomaly tracing;

[0044] Figure 4 Solve the logic diagram for the MILP scheduling module;

[0045] Figure 5 This is a topology and measurement point diagram of the rooftop photovoltaic + energy storage park in Example 1;

[0046] Figure 6 The curve showing the comparison between load and carbon emission predictions on September 15, 2025, in Example 1.

[0047] Figure 7 The figures show the power purchase and SOC curves before and after scheduling optimization in Example 1. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0049] 1. System Hardware Configuration

[0050] 1.1 Perception Layer Hardware

[0051] Power acquisition: A three-phase multi-function rail meter (accuracy class 0.5S) is used, which supports RS-485 / Modbus-RTU protocol, sampling period of 1 minute, and can be expanded to IEC 61850 GOOSE protocol to adapt to grid-side data access;

[0052] Cold / Heat Acquisition: Adopts ultrasonic cold / heat meter, supports 4-20mA + pulse dual output, pipe diameter covers DN50-DN200, conforms to EN1434 standard, and ensures the accuracy of cold / heat acquisition;

[0053] Gas / hydrogen acquisition: A thermal mass flow meter (accuracy class 0.5) is used, which supports 4-20mA+Modbus-TCP protocol and is suitable for working conditions with hydrogen purity ≥99.97%.

[0054] Meteorological data acquisition: An integrated micro-weather station is used, capable of collecting wind speed (accuracy ±0.1m / s) and irradiance (accuracy ±1W / m²). 2 Temperature (accuracy ±0.2℃) is wirelessly uploaded via the LoRaWAN protocol;

[0055] Carbon emission factor acquisition: The edge gateway has a built-in MQTT subscription module to subscribe to the Southern Power Grid's 1-hour marginal emission factor API in real time, cache it locally for 24 hours, and automatically fill in the missing data after network outage.

[0056] 1.2 Edge Layer Hardware and Software

[0057] Hardware configuration: It adopts an ARM Cortex-A72 quad-core processor (1.5GHz), equipped with 2GB LPDDR4 memory, 32GB eMMC storage, dual network ports (WAN / LAN), and supports 4G / 5G module expansion to meet the network needs of different scenarios;

[0058] Software configuration: Runs on Ubuntu Core 22.04 operating system and Docker 24.0 containerized environment, with a built-in edge data quality engine. Specific functions include:

[0059] ① Missing value handling: When the proportion of missing values ​​is <1%, linear imputation is used; when the proportion of missing values ​​is ≥1%, a lightweight LSTM imputation model is triggered (distilled and compressed to 2MB) to ensure data integrity.

[0060] ② Outlier handling: The 3σ criterion and the isolated forest algorithm are used to mark outliers in the collected data. After marking, the data is uploaded to the platform layer with an anomaly label to facilitate subsequent source tracing.

[0061] Resume interrupted transmission: The system uses an SQLite database to store the raw data packets from the most recent 48 hours. The transmission will resume automatically after the network is restored. MD5 checksums are used to ensure that the data is not lost or duplicated.

[0062] 1.3 Platform Layer Hardware and Software Deployment

[0063] Deployment environment: Utilizes a K8s 1.28 cluster, deployed with one click via Helm, supports microservice horizontal scaling (HPA), with single-service CPU usage ≤500MB and memory usage ≤512MB, meeting high concurrency requirements;

[0064] Service decomposition: The platform layer is divided into 6 microservices, namely ai-load (AI load prediction), ai-carbon (carbon emission inversion), ai-shap (anomaly tracing), ai-milp (low-carbon scheduling), api-gateway (API gateway), and web-portal (visualization service). Each service is deployed independently and works collaboratively.

[0065] Data storage and transmission:

[0066] ① Data bus: Kafka 3.5 is used. Topics are divided according to "energy medium + time granularity". Data is retained for 7 days and the compression rate is ≥70% to ensure efficient data transmission.

[0067] ② Time-series database: TimescaleDB 2.11 is used, with (device_id, time) as the partition key. The data write rate is ≥500,000 points / s and the query latency P99 is <200ms, which meets the requirements of massive time-series data storage and fast query.

[0068] Model repository: Utilizes MLflow 2.8, supports AI model version control, A / B testing, and model rollback, and can roll back to any historical version, facilitating model iteration and optimization.

[0069] 1.4 Application Layer Hardware Requirements

[0070] Visualization terminal: Supports access from PC and tablet. Minimum configuration: CPU i5 or above, memory 8GB or above, graphics card supporting 1080P or above resolution to ensure smooth display of the visual cockpit.

[0071] - Interface adaptation: Supports Ethernet and 4G / 5G network access, third-party API interfaces support JWT authentication, return data in JSON format, and have a concurrent processing capacity of >1000 TPS.

[0072] The table below shows the four-level architecture and data flow of the system of this invention.

[0073] 2. Core Algorithm Flow

[0074] 2.1 AI Load Forecasting Module Algorithm Flow

[0075] This module adopts an algorithm architecture of "parallel base model + meta-learning fusion", see appendix. Figure 2 The specific steps are as follows:

[0076] Step 1: Feature Input and Preprocessing

[0077] Energy domain characteristics: Real-time power L_i(t) of five energy media (electricity, cooling, heating, gas, and hydrogen) is collected, and historical 7-day load lag, wavelet packet entropy (reflecting load fluctuation characteristics), trend term, periodic term, and residual term after STL decomposition are extracted.

[0078] Carbon domain characteristics: Obtain the dynamic carbon emission factor EF_i(t) for each energy medium and calculate the equivalent carbon load:

[0079]

[0080] Event characteristics: By mining events such as production plan adjustments, equipment maintenance, and sudden weather changes through knowledge graph, the event confidence vector Evt (dimension k=32, confidence value 0-1) is output to quantify the impact of the event on the load;

[0081] Feature preprocessing: Standardize all input features (normalize to the [0,1] interval), remove redundant features, and improve model training efficiency.

[0082] Step 2: Parallel base model training

[0083] Three base models are trained in parallel, each leveraging its strengths. Specific parameter configurations are as follows:

[0084] XGBoost model: n_estimators=800, max_depth=7, learning_rate=0.05, negative sampling ratio 1:1, trained with a 4-core CPU, training time ≤3min, mainly used to capture non-linear relationships between features;

[0085] TCN model: dilation=[1,2,4,8,16], kernel_size=3, dropout=0.05, receptive field≥24h, number of parameters<200k, mainly used to capture long-term dependencies in load time series;

[0086] GRU model: hidden_size=64, number of layers=2, dropout=0.1, accelerated by CuDNN, inference latency<50ms, mainly used to improve the real-time inference speed of the model.

[0087] Step 3: Fusion of LightGBM meta-learners

[0088] Fusion input: The outputs of the three parallel base models, the event vector Evt (32-dimensional), the equivalent carbon load LCE(t), and the load prediction residual ε are used as inputs to the LightGBM model. The total input dimension = 3 + 32 + 1 + 1 = 37.

[0089] Model parameters: n_estimators=500, max_depth=5, feature_fraction=0.8, using a five-fold cross-validation early stopping strategy to avoid model overfitting;

[0090] Fusion effect: The MAPE of the day-ahead load forecast after fusion is reduced by ≥1.2 percentage points compared with the best single model (XGBoost), and finally the MAPE is ≤3.6%.

[0091] Step 4: Online Correction and Incremental Training

[0092] Scrolling window adjustment: When the prediction error > 5%, the scrolling window is automatically adjusted. The window length W is calculated using the following formula: , where W∈[7,28] days, to ensure that the window length matches the prediction accuracy;

[0093] Incremental training: A warm-start strategy is adopted, which only updates the last 20% of the decision trees in the LightGBM model. There is no need to retrain the entire model. The training time is ≤90s, ensuring that the system outputs high-precision load prediction results 24 / 7.

[0094] 2.2 Algorithm Flow of Carbon Emission Inversion Module

[0095] This module, based on dynamic carbon emission factors and event corrections, enables accurate carbon emission calculation. (See appendix.) Figure 3 Specific process:

[0096] Core accounting formula:

[0097]

[0098] in:

[0099] ① : The predicted load of the i-th energy medium output by the AI ​​load prediction module (the real-time load collected by the perception layer is used during real-time scheduling).

[0100] ② Dynamic carbon emission factor, sources include the 1-hour marginal emission factor of China Southern Power Grid (grid side) and self-owned photovoltaic power generation percentage. (in ), Supplier's monthly measured values ​​(purchased steam / hydrogen);

[0101] Supplier's monthly measured values ​​(purchased steam / hydrogen);

[0102] ③ η_i: Energy conversion efficiency, preset according to equipment parameters, and can be calibrated periodically using historical operating data;

[0103] ④ t: Event correction item, calculated using the following formula: , This represents the event's impact coefficient.

[0104] Accuracy verification: Calculate the calculated carbon emission values ​​with the actual monitored values ​​(if any) every hour, calculate the calculation deviation, and automatically trigger the anomaly tracing module when the deviation exceeds ±3.3%, while simultaneously calibrating the event impact coefficient. This ensures that the accuracy of accounting remains stable over the long term.

[0105] 2.3 Algorithm Flow of Self-Explanatory Anomaly Tracing Module

[0106] This module uses the SHAP value interpretation algorithm to quickly locate carbon emission anomalies. The specific process is as follows:

[0107] Triggering condition: Calculate the actual value of carbon emissions With accounting value The deviation, when | – When |> α (α = 1.5σ, where σ is the standard deviation of the historical 30-day accounting deviation), the anomaly tracing will be automatically triggered;

[0108] SHAP value calculation: The TreeSHAP algorithm is used to calculate the feature contribution values ​​output by the XGBoost and LightGBM models, and the DeepSHAP algorithm is used to calculate the feature contribution values ​​output by the GRU and TCN models. The calculation time is <1min.

[0109] Anomaly tracing output: Sort the contribution values ​​of all features and output the top-3 high-contribution anomaly features (such as "abnormal electricity purchase load due to electricity price fluctuations", "untimely update of carbon emission factors", "increased energy loss due to equipment failure") and their corresponding contribution values. At the same time, a tracing report is generated to help operation and maintenance personnel quickly find the root cause of the anomaly.

[0110] 2.4 MILP Low-Carbon Scheduling Module Algorithm Flow

[0111] This module is based on mixed-integer linear programming to achieve 15-minute rolling low-carbon scheduling. (See attached file.) Figure 4 Specific process:

[0112] Definition of decision variables:

[0113] Energy storage charging and discharging power ,

[0114] Chiller unit start / stop u_chiller(t)∈{0,1}

[0115] Solar curtailment ≥0

[0116] Objective function construction: The objective function is as follows, aiming to minimize the total cost of electricity purchase, carbon emission cost, and energy storage lifetime penalty:

[0117]

[0118] In the formula: Let t be the power purchased at time t. Let t be the electricity purchase price. C(t) represents the local carbon trading price (yuan / tCO2), C(t) represents the total carbon emissions at time t, and κ represents the energy storage cycle depreciation factor (yuan / kWh). = - .

[0119] Constraint settings:

[0120] ① Power balance constraints: To ensure a balance between power supply and power load;

[0121] ② Energy storage lifetime constraints: ,in For the rated capacity of energy storage, =365×2 (designed for two charging and two discharging cycles per day);

[0122] ③ Constraints on electricity purchase demand: ≤ ,in The upper limit for electricity purchase demand is set, and any excess is penalized with three times the electricity price to control electricity purchase costs.

[0123] ④ Equipment operation constraints: chiller start-up and shutdown interval ≥ 30 min, energy storage charging and discharging power ≤ rated power, photovoltaic output ≤ maximum power generation.

[0124] Solution and Output: The OR-Tools 9.7 solver is used, with MIPGap=0.5% set. The solver is rolled once every 15 minutes, and the solution time is ≤30 seconds. The output includes the power purchase, energy storage charging and discharging power, chiller start and stop status and other scheduling strategies at time t. These are then pushed to the application layer visualization dashboard for operation and maintenance personnel to execute.

[0125] 3. Interface and reproducibility design (meeting the "manufacturability or usability" requirement for patent utility)

[0126] 3.1 Southbound Interface (Device Access Interface)

[0127] Supports five mainstream industrial protocols: Modbus-RTU / TCP, IEC104, MQTT, BACnet, and OPC UA, adaptable to different types of data acquisition and energy equipment. Specific design details:

[0128] Protocol conversion: The edge layer has a built-in protocol conversion module that can convert the private protocols of different devices into the standardized Modbus-TCP protocol and upload them to the platform layer;

[0129] Zero-code configuration: When connecting a device, no coding is required. Simply fill in the device name, register address, range, coefficient and other parameters through the configuration interface (the configuration table is in YAML format) to complete the device connection, adapting to the diverse needs of devices in multiple scenarios.

[0130] 3.2 Northbound Interface (Third-Party Platform Integration Interface)

[0131] Provides RESTful API and message push interface, supports integration with third-party systems, and the specific design is as follows:

[0132] RESTful API interface: The interface path includes / v1 / load (load forecast data), / v1 / carbon (carbon emission data), and / v1 / alarm (abnormal alarm data). It adopts the JWT authentication mechanism to ensure interface security, returns data in JSON format, and has a concurrent processing capacity of >1000 TPS.

[0133] Message push interface: Supports three message buses: Kafka, RabbitMQ, and ActiveMQ. The topic naming convention is "park.scene.device.metric" (park.scene.device.metric), which facilitates data subscription by third-party systems.

[0134] 3.3 One-click deployment and zero-code copy design

[0135] To enable rapid replication across multiple scenarios, a Helm one-click deployment script and a zero-code configuration tool were designed. The specific implementation is as follows:

[0136] One-click deployment script: Helm Chart contains 8 YAML configuration files including ConfigMap, Deployment, Service, HPA, and CronJob, covering the deployment configuration of all microservices, databases, and data buses at the platform layer, and can complete the deployment of a new campus system within 30 minutes;

[0137] Offline deployment support: Provides an offline installer package for docker-compose, with an image size of <2GB, including offline images of all microservices and databases, suitable for deployment in scenarios without a public network environment;

[0138] Zero-code scenario configuration: When adding a new scenario, the system can be deployed in the new scenario by selecting the energy medium type, equipment parameters, scheduling target, etc. through the application layer configuration interface. No customized coding is required, and the configuration time is ≤2 days.

[0139] 4. Application Cases

[0140] The following two application examples in different scenarios demonstrate the technical effectiveness, practicality, and reproducibility of the present invention. These examples are only one of the many application scenarios and do not limit the scope of application of the present invention.

[0141] 4.1 Case Study 1: A Semiconductor Industrial Park in Shenzhen (Rooftop Solar PV + Energy Storage Scenario)

[0142] 4.1.1 Basic Information

[0143] The semiconductor park covers an area of ​​approximately 50,000 square meters and consumes approximately 2 GWh of electricity annually. It is equipped with a 280 kW rooftop photovoltaic system and a 100 kW / 215 kWh energy storage system. The energy mediums include electricity, steam, and compressed air (corresponding to the "gas" medium in this invention). It is necessary to achieve efficient energy utilization and precise carbon emission control.

[0144] 4.1.2 System Deployment

[0145] Sensing layer: 12 new three-phase multi-functional rail meters, 4 heat meters, 2 gas meters, and 1 integrated micro weather station have been added, covering the three core energy media of electricity, heat and gas, with a sampling cycle of 1 minute.

[0146] Edge layer: Deploy one ARM Cortex-A72 quad-core edge gateway, locally cache 48h of data, support 4G module expansion, and ensure that data is not lost in the event of network outage;

[0147] Platform layer: It adopts K8s cluster deployment, deploys 6 microservices with one click through Helm, and configures TimescaleDB time series database and Kafka data bus;

[0148] Application layer: Deploy a visual dashboard, configure automatic carbon emission report generation function, and connect to the park's existing ERP system.

[0149] 4.1.3 Operational Data (September 1, 2025 to October 31, 2025, a total of 61 days)

[0150] Load forecasting accuracy: The day-ahead load forecasting MAPE is 3.1%, which is far lower than the target of ≤3.6% set by this invention. Compared with the traditional LSTM model (MAPE=9.4%), the accuracy is improved by 67%.

[0151] Carbon emission accounting accuracy: The carbon emission accounting deviation is ±2.8%, which is lower than the target of ≤±3.3% set by this invention. Compared with traditional manual accounting (deviation ±12.1%), the accuracy is improved by 77%.

[0152] Anomaly tracing results: The load prediction error for extreme days (Mid-Autumn Festival + equipment maintenance) was 5.1%, which is lower than the 6% assessment line. The anomaly tracing time was 45 seconds, and the two high-contribution anomaly features of "production plan adjustment" and "sudden weather change" were successfully located.

[0153] Scheduling optimization results: The MILP low-carbon scheduling module solves the problem in 15 minutes, with peak energy storage discharge of 95kW and valley charging of 85kW, reducing peak electricity purchase by 1.5MWh and carbon emissions by 1.1t per day.

[0154] Economic benefits and carbon emission reduction effects: Annual peak-valley arbitrage and demand management revenue increased by RMB 187,000, and annual economic benefits increased by 23% (higher than the ≥20% set by this invention); comprehensive carbon emissions decreased by 9.2% (higher than the ≥8% set by this invention); energy storage cycle count decreased by 11%, and lifespan was extended by 1.2 years.

[0155] 4.2 Case Study 2: A Data Center in Shenzhen (High Power Load Scenario)

[0156] 4.2.1 Basic Information

[0157] The data center has an IT load of 1.5MW and an annual electricity consumption of approximately 1.3GWh. The main energy sources are electricity and chilled water (air conditioning system). It is necessary to achieve accurate power load forecasting, carbon emission control, and cost optimization.

[0158] 4.2.2 System Deployment and Replicability Verification

[0159] Using the same system container image from Case 1, deployment was performed via zero-code configuration. The specific process is as follows:

[0160] Sensing layer: Add 8 power data acquisition devices, 3 cold water flow data acquisition devices, and configure 1 weather station with a sampling cycle of 1 minute;

[0161] Edge layer and platform layer: Reuse the container image from Case 1 and deploy it with one click using Helm, only modifying device parameters, energy medium type and other configurations;

[0162] Deployment cycle: The total configuration time from equipment installation to system launch is 1.5 days, which is lower than the target of ≤2 days set by this invention.

[0163] 4.2.3 Running Results

[0164] Load forecast accuracy: The day-ahead power load forecast MAPE is 3.4%, which meets the target of ≤3.6% set by this invention;

[0165] Carbon emission accounting accuracy: The carbon emission accounting deviation is ±3.0%, which meets the target of ≤±3.3% set by this invention;

[0166] Economic benefits and carbon emission reduction: Annual electricity purchase cost reduced by RMB 168,000, annual economic benefits increased by 21%; comprehensive carbon emissions reduced by 8.5%, all of which meet the targets set by this invention.

[0167] 4.2.4 Reproducibility Conclusion

[0168] The same system container image can be quickly deployed to two different scenarios, semiconductor parks and data centers, through zero-code configuration, with a deployment cycle of ≤2 days. Both deployments can meet the technical indicators set by this invention, proving that this invention has good universality and replicability and can be widely applied to various integrated energy scenarios.

[0169] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-based intelligent energy and carbon management system, employing a four-level architecture of "sensing-edge-platform-application," characterized in that... include: a) Sensing layer, used for synchronous acquisition of multi-energy flow data, covering at least five energy media: electricity, cooling, heating, gas and hydrogen, with a sampling period of ≤1min, and simultaneously acquiring data related to meteorology, electricity price, production plan, maintenance order and dynamic carbon emission factor; b) Edge layer, with built-in edge computing gateway, to complete the functions of cleaning, protocol conversion, local caching and breakpoint resume of collected data, where the local caching time is ≥48h; c) At the platform layer, a microservice containerized deployment approach is adopted, including at least: The AI ​​load forecasting module is used to output load forecast curves at multiple time scales, including day-ahead, intraday, and real-time forecasts. The mean absolute percentage error (MAPE) of the day-ahead load forecast is ≤3.6%. The carbon emission inversion module is used to calculate the carbon emission curve based on the predicted load output by the AI ​​load prediction module and the dynamic carbon emission factor collected by the sensing layer. Its carbon emission calculation deviation is ≤±3.3%. The self-explaining anomaly tracing module is used to locate the top-3 high-contribution anomalies when the deviation between the actual and calculated carbon emission values ​​exceeds a preset threshold, with an anomaly tracing time ≤ 1 minute. The MILP low-carbon scheduling module aims to minimize the cost of electricity purchase, carbon emission cost, and energy storage lifetime penalty. It generates low-carbon scheduling strategies using a 15-minute rolling solution method, with a solution time of ≤30 seconds. d) Application layer, used to provide a visual cockpit, automatic generation of carbon emission reports and third-party API interfaces, with a configuration time of ≤2 days for new scenario access.

2. The AI-based intelligent energy and carbon management system according to claim 1, characterized in that: The AI ​​load prediction module employs a parallel base model of "XGBoost+TCN+GRU" and fuses the knowledge graph event confidence vector Evt and equivalent carbon load through a LightGBM meta-learner. Where the dimension k of the Evt vector is ≤32, the formula for calculating the equivalent carbon load LCE(t) is: In the formula For the real-time load of the i-th energy medium, is the dynamic carbon emission factor for the i-th energy medium.

3. The AI-based intelligent energy and carbon management system according to claim 1, characterized in that: The carbon emission inversion module uses the following calculation formula to calculate carbon emissions: In the formula Let be the total carbon emissions at time t. Let be the energy conversion efficiency of the i-th energy medium. This is an event correction item used to correct the impact of various unforeseen events on carbon emission accounting results.

4. The AI-based intelligent energy and carbon management system according to claim 1, characterized in that: The MILP low-carbon scheduling module linearizes the energy storage cycle life loss, with a linear fit goodness of ≥0.95, and supports a two-charge and two-discharge scheduling strategy for energy storage, with a scheduling time granularity of ≤15min.

5. The AI-based intelligent energy and carbon management system according to any one of claims 1-4, characterized in that, The platform layer supports horizontal scaling, with a single-node fault recovery time (RTO) of ≤5 min, meeting the requirement for continuous and stable operation 24 / 7.

6. An AI-based digital energy and carbon management method, characterized in that, Using the AI-based intelligent energy and carbon management system as described in any one of claims 1-5, the following steps are completed sequentially: S1: The sensing layer completes the synchronous acquisition of multi-energy medium load and related data, and uploads it to the edge layer; S2: The edge layer cleans and converts the collected data according to the protocol, caches it locally and uploads it synchronously to the platform layer. Local caching is completed when the network is disconnected, and interrupted transmission is resumed after the network is restored. S3: The platform layer completes multi-timescale load forecasting through the AI ​​load forecasting module, completes carbon emission accounting through the carbon emission inversion module, completes carbon emission anomaly location through the self-explaining anomaly tracing module, and generates rolling low-carbon scheduling strategies through the MILP low-carbon scheduling module. S4: The application layer receives data and strategies output from the platform layer, displays them in a visual cockpit, generates carbon emission reports, or synchronizes them to other systems via third-party API interfaces.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the AI-based intelligent energy and carbon management method as described in claim 6.

Images