District energy management system and method based on synergistic intelligence and dynamic adaptation

Abstract

This invention discloses a regional energy management system and method based on collaborative intelligence and dynamic adaptation. The system operates in a two-level optimization architecture consisting of a regional command layer and a single-building model layer. Privacy-preserving collaborative intelligence sharing is achieved through a federated learning mechanism. The regional command layer receives and aggregates model updates to refine a global energy optimization model, and adaptively adjusts the regional optimization strategy using this model. The output of the single-building model layer continuously performs closed-loop calibration and updates to the digital twin network. Based on real-time detected trigger events or predicted conditions, the regional command layer dynamically adjusts the control granularity of the single-building model layer or temporarily authorizes intermediate cluster-level models. This aims to address the challenges of data privacy, computational complexity, and insufficient system resilience in the management of large-scale distributed energy systems.

Description

Technical Field

[0001] This invention belongs to the field of energy management, and specifically relates to a regional energy management system and method based on collaborative intelligence and dynamic adaptation. Background Technology

[0002] Traditional regional energy management systems typically rely on a single centralized optimizer for macro-level decision-making and overall control. This architecture centralizes all energy management tasks into a single global entity, aiming to maximize efficiency through global optimization of the entire regional energy system.

[0003] However, when faced with modern large-scale, highly heterogeneous, and data-sensitive distributed energy systems, this traditional energy management paradigm with a single centralized optimizer faces the following key technical challenges and limitations: 1. High computational complexity and insufficient scalability The existence of regional command layers is a fundamental response to the challenges faced by a single centralized optimizer in traditional energy management.

[0004] • Computational Insufficiency: As the system scales up and the number of integrated distributed energy assets (DERs) and individual buildings increases, a single, global optimizer often becomes computationally inadequate when dealing with large-scale, heterogeneous, and data-sensitive distributed systems. This is often referred to as the NP-Hard problem.

[0005] • Centralized computational burden: Centralized systems concentrate all computational load on a central server. As the scale increases, this significantly increases the computational latency and network bandwidth requirements of the central entity, thereby limiting the system's scalability and real-time responsiveness.

[0006] 2. Data privacy and security risks In order for a single centralized optimizer to perform global optimization accurately, it needs access to the details of all underlying individual buildings.

[0007] • Sensitive data leakage: This traditional model requires individual building AI models to share raw or semi-raw operational data (e.g., indoor temperature, occupancy rate, HVAC system performance, local renewable energy generation, etc.) directly with the central regional model.

[0008] • Privacy Barriers: This practice of directly sharing sensitive data with a central entity raises significant data privacy, security, and computational overhead issues. Data privacy has become a major obstacle to achieving regional collaborative optimization.

[0009] 3. Lack of effective mechanisms to address local differences A single centralized optimizer struggles to balance global objectives with localized needs.

[0010] • Rigid instructions: Traditional centralized optimizers struggle to effectively guide the micro-level refinement of local execution. They typically issue rigid instructions, preventing local models from autonomously selecting the execution strategy that best suits their local needs based on their comfort level, energy consumption, and operational constraints.

[0011] • Inefficiency: A single optimizer is inefficient when dealing with large-scale, heterogeneous, and data-sensitive distributed systems.

[0012] Therefore, there is an urgent need for a new architecture to overcome the inherent shortcomings of a single centralized optimizer. This architecture must maintain consistency of global objectives, significantly avoid the enormous computational overhead of global optimization, protect local data privacy, and simultaneously achieve refined and autonomous management of underlying energy assets. This is precisely why this invention adopts a two-tier optimization architecture of "regional command, individual execution" to replace or supplement the traditional centralized management model. Summary of the Invention

[0013] To address the aforementioned technical problems, this invention proposes a regional energy management system based on collaborative intelligence and dynamic adaptation, specifically comprising: It operates in a two-level optimized architecture, which includes a regional command layer and multiple individual building model layers; Privacy-preserving collaborative intelligent sharing is achieved through a federated learning mechanism. The federated learning includes: the multiple individual building model layers independently training local energy optimization sub-models using their local real-time operating data; the multiple individual building models generating anonymized and encrypted model updates based on the local energy optimization sub-models without transmitting the local real-time operating data; and the regional command layer receiving and aggregating the model updates to refine a global energy optimization model, and adaptively adjusting the regional optimization strategy using the global energy optimization model. By continuously calibrating and updating the digital twin network through the output of the individual building model layer, a dynamic virtual representation of the building's energy and behavior can be achieved, enabling proactive predictive management of the building. The regional command layer dynamically adjusts the control granularity of the individual building model layer or temporarily authorizes the intermediate cluster-level model based on real-time detected trigger events or predicted conditions. The intermediate cluster-level model is located between the regional command layer and the individual building model layer.

[0014] Specifically, the control granularity of the individual building model layer includes temporal granularity, spatial granularity, or operational granularity; The dynamic adjustment control granularity is achieved through an enhanced dynamic adaptive optimization mechanism, which includes: The regional command layer uses an autoencoder to calculate the system state vector. Reconstruction error ; The reconstruction error The adaptive dynamic programming model, used as the stage cost input, undergoes model-free iterative optimization to generate a robust risk assessment. With optimal risk control ; based on Change in reconstruction error Adjust the control interval continuously according to the following formula. Achieve time-domain granularity refinement and sub-minute response:

[0015] in Indicates the reference control interval. Indicates the risk sensitivity coefficient. This represents the sensitivity coefficient to changes in error.

[0016] Specifically, the dynamic adjustment control granularity is further refined in terms of spatial granularity and adjusted in terms of operational granularity through a hierarchical deep reinforcement learning agent, specifically including: The regional command level employs a layered strategic agent, biased based on contextual and weather perception. Select the optimal topology configuration To determine the spatial optimization scale ; Enabling hierarchical tactical agents to optimize continuous action vectors under multi-energy coupling constraints. The vector includes at least the energy storage power. Thermal storage power or cold storage power ; The tactical agent maximizes policy entropy. Ensure exploration efficiency and strategy robustness to achieve component-level fine-grained control over specific energy assets.

[0017] Specifically, the cluster-level model of the temporary authorization intermediate includes: When there are significant internal interactions and shared energy assets in a sub-region, the regional command layer intelligently activates or temporarily authorizes the intermediate cluster-level model. The intermediate cluster-level model receives the aggregation targets issued by the regional command layer. And within their jurisdiction, they shall perform localized demand response coordination or shared energy storage optimization; By delegating fine-grained optimization tasks to the intermediate cluster-level model, global computational overhead is avoided, significantly improving system resilience and rapid response capabilities.

[0018] The present invention also discloses a regional energy management method based on collaborative intelligence and dynamic adaptation, which operates in a two-level optimization architecture, including a regional command layer and multiple individual building model layers; Privacy-preserving collaborative intelligent sharing is achieved through a federated learning mechanism. The federated learning includes: the multiple individual building model layers independently training local energy optimization sub-models using their local real-time operating data; the multiple individual building models generating anonymized and encrypted model updates based on the local energy optimization sub-models without transmitting the local real-time operating data; and the regional command layer receiving and aggregating the model updates to refine a global energy optimization model, and adaptively adjusting the regional optimization strategy using the global energy optimization model. By continuously calibrating and updating the digital twin network through the output of the individual building model layer, a dynamic virtual representation of the building's energy and behavior can be achieved, enabling proactive predictive management of the building. The regional command layer dynamically adjusts the control granularity of the individual building model layer or temporarily authorizes the intermediate cluster-level model based on real-time detected trigger events or predicted conditions. The intermediate cluster-level model is located between the regional command layer and the individual building model layer.

[0019] The method of this invention integrates three key technical mechanisms: 1. Federated Learning Mechanism: This mechanism enables multiple individual building AI models to independently train energy optimization sub-models locally using their private data, and only transmits anonymized and encrypted model updates, rather than the original sensitive data, to the regional command layer. The regional command layer then uses algorithms to refine the global energy optimization model, thereby achieving decentralized computational load and collaborative intelligent sharing while protecting local data privacy.

[0020] 2. Digital Twin Network Strategic Integration: Introducing a digital twin network for individual buildings. The outputs of each individual building's AI model (such as operational control commands and predictive insights, including RUL and component degradation predictions) are continuously fed into the BDT to calibrate and update the virtual representation, thereby elevating the system from reactive operation to proactive predictive management and enabling regional predictive maintenance optimization and enhanced resilience planning.

[0021] 3. Event-driven dynamic adaptive control: Based on real-time detected trigger events or predicted conditions, the regional command layer uses advanced AI algorithms (such as hierarchical DRL, ADP fusion, and autoencoder anomaly detection) to dynamically adjust the control granularity (including temporal, spatial, or operational granularity) or temporarily authorize intermediate cluster-level AI models.

[0022] Through the above mechanism, this method achieves decentralized computing load, avoids global computing overhead, and can achieve sub-minute response to events such as power grid anomalies and local congestion, thereby significantly enhancing the system's resilience and rapid response capability. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the collaborative intelligent and dynamically adaptive regional energy management system proposed in this invention. Detailed Implementation The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to specific embodiments.

[0025] This invention proposes a regional energy management system based on collaborative intelligence and dynamic adaptation, such as... Figure 1 As shown, the system operates within a two-level optimized architecture, which includes a regional command layer and multiple individual building model layers. The regional command level is responsible for macro-management, overall optimization, and setting global boundary conditions and target parameters.

[0026] The regional command level is responsible for refining the overall energy optimization model. It also utilizes a global model to generate strategic regional optimization outputs, including carbon intensity targets. and / or demand response instructions 3. Context awareness and dynamic adjustment: Equipped with advanced... (such as distributed reinforcement learning) (Or an anomaly detection algorithm) to continuously monitor multi-dimensional contextual factors and, based on real-time detected triggering events or predictive conditions. Dynamically adjust the control granularity of the lower layers. How can the regional command layer dynamically adjust its control granularity over the lower-level individual building AI model layer based on real-time detected trigger events or predicted conditions, thereby significantly enhancing the system's resilience and rapid response capability?

[0027] Definition and mechanism of control granularity Control granularity refers to the level of detail or resolution by which regional command levels monitor, model, and control the energy system; its dynamic adjustment is based on the results of risk assessments. drive.

[0028] 1. Contextual awareness and risk assessment Regional command level continuous monitoring system state vector ,

[0029] 1. Robust Risk Assessment (AE+ADP Fusion): Reconstruction Mechanism (RM) uses an autoencoder (AE) and AI algorithms (such as an autoencoder AE) to calculate the reconstruction error.

[0030] The robust risk assessment R (e.g., R∈0,1,2) is generated through model-free iterative optimization using adaptive dynamic programming (ADP). R does not depend on static thresholds but is based on the minimization of long-term cumulative costs determined by ADP.

[0031]

[0032] The regional command level (RM) continuously refines the global energy optimization model using federated learning. This ensures that strategic decisions, such as risk weights γ, are based on collective intelligence. The system employs a hierarchical DRL agent structure, where the regional command layer strategic agent (PPO) is responsible for selecting the topology configuration c (spatial granularity, discrete actions), and the tactical agent (SAC) is responsible for optimizing the continuous action vector a (operational granularity, continuous actions).

[0033] An enhanced layered DRL architecture is adopted, which integrates regional command layer strategic agency. Used for strategic decision-making and tactical agency (SAC) for tactical execution to address real-time triggering events and complex multi-energy coupling constraints in regional energy systems.

[0034] Regional Command Layer Strategic Agent (PPO): Determining Topology Configuration and Spatial Optimization Scale. The Strategic Agent resides at the Regional Command Layer (RM) and is responsible for determining the scale based on the extended system state vector. Making macro-level decisions, i.e., selecting the optimal topology configuration. This implicitly determines the spatial optimization scale. .

[0035] Input status of strategic agent It is an extended state vector, including real-time running state. Power grid condition prediction External factors Weather / Environmental Factors (e.g., humidity and temperature forecasting) and multi-energy loads (e.g., hot / cold load).

[0036]

[0037] Strategic agency The Proximal Policy Optimization (PPO) algorithm is used for policy optimization. Strategic Policy The goal is to select a discrete topology configuration. .

[0038] Topology configuration This represents different combinations of control strategies or authorization levels, and directly determines the scale of spatial optimization. .For example: : (Global goals issued).

[0039] : (Temporary authorization cluster level) Model Launch is applicable to complex sub-regions with shared energy assets, such as university campuses.

[0040] : (Issuing highly specific boundary conditions to individual buildings within a specific sub-region affected by the boundary conditions).

[0041] Strategic agents integrate weather perception biases To guide decision-making. If extreme weather (such as high temperatures) is predicted. We will tend to choose topology configurations that can effectively utilize cooling assets.

[0042]

[0043] in It is a state go through The representation after the proxy network, Based on weather indicators and bias vector The calculated bias.

[0044] Layered Tactical Agents (SAC) are used to optimize continuous action vectors. The tactical agent receives the configuration selected by the strategic agent. As part of its extended state, and responsible for optimizing the continuous action vector This enables fine-grained control at the component level.

[0045] Tactical agent adoption The SoftActor-Critic algorithm supports continuous action spaces. (Continuous action vectors) This includes at least the critical storage power in multi-energy coupled systems:

[0046] in: The charging and discharging power (continuous value, e.g.) of the battery energy storage system ).

[0047] Power of the thermal storage system (thermal energy injection / extraction rate, continuous value).

[0048] Power of the cold storage system (cold energy injection / extraction rate, continuous value).

[0049] Action output is usually achieved through Function normalization to Range, to handle continuous space.

[0050]

[0051] The core of tactical agency lies in maximizing policy entropy. To ensure exploration efficiency and policy robustness, the optimization objective of SAC is to maximize the sum of cumulative reward and policy entropy.

[0052] in It's a temperature parameter used to balance rewards. and policy entropy .

[0053] The Q and V functions of SAC also explicitly include entropy terms:

[0054] By maximizing entropy, tactical agents avoid over-reliance on a single optimal path, maintaining policy robustness even in unforeseen situations (such as intermittent renewable energy or local sensor failures) and continuing to effectively explore the action space, thereby improving energy savings.

[0055] Furthermore, multi-energy coupling constraints and component-level refined control tactical agents in optimization At the same time, the multi-energy coupling constraints in the system must be observed.

[0056] In a cluster containing a combined electricity, heating, and cooling system, the following is an example of the constraints: For example, combined heat and power (CHP) units or The output power of a heat recovery heat pump must meet its coefficient of performance. Requirements for (performance coefficient).

[0057]

[0058] Lithium batteries thermal storage and cold storage The charging and discharging must comply with its (State of charge) dynamics and capacity limitations.

[0059]

[0060] In the process of component-level fine-grained control, the tactical agent outputs a continuous power vector. This enables fine-grained control of specific energy assets at the component level. For example, during peak cooling load periods, the tactical agent can perform the following actions: Continuous adjustment: Adjusting the chiller unit Pattern from Adjust to (Refine parameters), and at the same time... Adjust to maximum cooling injection power (Continuous action).

[0061] SOC Collaboration: If There is still a margin; the agent may output a negative value. Value (discharge) to support To meet the electricity demand and achieve power-cooling synergy.

[0062] The following example illustrates a cluster-level dynamic adjustment triggered by localized high temperatures. It is assumed that the regional command level detects a localized commercial cluster. Traffic congestion is imminent, and weather forecasts indicate... The ambient temperature is expected to reach an extreme peak (leading to a surge in air conditioning load).

[0063]

[0064] Context awareness: The regional command level detected... Temperature index in Higher, calculated This bias encourages the selection of those who can utilize... The topology configuration of complex multi-energy assets (especially central refrigeration and combined heat and power).

[0065] Configuration Options: Strategic Agent according to choose (Authorized cluster level) Model ).

[0066] Spatial scale determination: Spatial optimization scale Dynamically shrink to The boundary.

[0067]

[0068] Authorization and target assignment: Regional command level to Cluster level Model (here) Internally contains Tactical agent) issues aggregated targets (For example, in the future) The total internal power load shall not exceed ).

[0069] Tactical maneuver optimization: of Tactical proxy Optimize the continuous action vector as the action boundary. .

[0070] Maximum Entropy Constraint: During the optimization process, the hierarchical tactical agent SAC maximizes the policy entropy. This is to ensure that its control strategy is robust enough to cope with the transient uncertainties of local load forecasting and renewable energy generation.

[0071] Multi-energy synergy and refined control: The proxy output is negative. (Discharge) to satisfy constraint.

[0072] At the same time, the agent will Set to the maximum value to extract energy from the chilled water storage, thereby reducing the power consumption of the real-time chiller unit.

[0073] if China has deployed The proxy output A negative value (releasing heat) enables heat load transfer and utilizes... The efficiency curve (i.e.) Coupling constraints enable optimal arbitration between electrical / thermal / cold components, achieving fine-grained control at the component level.

[0074] Specifically, the policy entropy regularization objective function (SAC)

[0075] in, This represents the optimal strategy, i.e., the final behavioral pattern that the tactical agent attempts to learn. This indicates the search for a strategy π that maximizes the expression on the right. This indicates the calculation of the expected value, typically across all states. The probability-weighted average of the actions. Indicates cumulative rewards. Represents policy π in a given state The policy entropy below, This represents the temperature parameter, which is an adjustable hyperparameter used to balance cumulative reward. The importance of the relationship between policy entropy (H(π)) and policy entropy (H(π)); The larger the value, the more the model tends to explore (the more random the strategy); The smaller the value, the more the model tends to utilize known optimal paths (the more certain the strategy), and the more the objective function will shift from the traditional goal of maximizing reward. With the goal of maximizing entropy By combining these approaches, an automatic balance between exploration and utilization is achieved.

[0076] ,

[0077] in, Parameters representing the strategic PPO agent network, Indicates topology configuration. This represents the expanded system state vector. This indicates the output of the PPO agent network. This indicates a weather perception bias. Indicates optional topology configuration. This represents the softmax function.

[0078] For cluster-level energy target matching constraints, the authorized cluster-level Model The macro-level objectives of the regional command must be met:

[0079] This hierarchical architecture utilizes regional command layer strategic agents. Solved the macro-situation judgment and resource allocation ( The strategic problem of ) is solved by using hierarchical tactical agent SAC to solve the tactical problem of continuous power adjustment and entropy maximization (robustness) under multi-energy constraints.

[0080] The regional command (RM) continuously monitors key performance indicators and the expanded system state vector across the entire region. Including real-time running status Power grid condition prediction External factors Weather / Environmental Factors (e.g., humidity and temperature forecasting) and multi-energy loads (e.g., hot / cold load).

[0081] When the regional command (RM) detects a predefined triggering event or predictive condition and assesses that the risk requires more granular spatial control, it will trigger dynamic adjustments.

[0082] Example of triggering an event: Localized energy supply and demand anomalies: for example, specific sub-regions (such as...) The load forecast suddenly increased dramatically.

[0083] Power grid stability issues: for example, local grid congestion or abnormal grid frequency / voltage.

[0084] Specific demand response command trigger: Emergency load reduction command issued by the grid operator for this sub-region.

[0085] Regional command (RM) utilizes advanced Algorithms (such as anomaly detection or distributed reinforcement learning) evaluate the situation and make decisions on activation or temporary authorization.

[0086] Assuming the regional command level (RM) predicts (A shared energy storage) and cogeneration The university campus clusters equipped with the equipment are about to face the risk of power grid congestion. RM made a strategic decision to grant cluster-level authorization to this sub-region. .

[0087]

[0088] Cluster level Once activated, the model acts as an intermediate layer, responsible for conducting more detailed and real-time energy arbitration within its jurisdiction.

[0089] Receive aggregated objectives issued by the regional command level The goal is to In the next optimization time domain Total load limit requirement within the area.

[0090]

[0091] Model predictive control can be used. or distributed reinforcement learning The framework is optimized with fine granularity.

[0092] The goal is to minimize total operating costs, comfort penalties, and penalties for not meeting regional objectives within the cluster.

[0093] in It is the current time step. It predicts the time domain. It's a direct penalty for comfort. It is a key element; it punishes behaviors that fail to meet regional objectives. Representing the energy cost at time step t, ensuring When enforcing local arbitration (such as peer-to-peer energy transactions), the aggregation targets issued by the regional command (RM) must be followed.

[0094] In the process of core fine-grained optimization and coordination, Within its jurisdiction, it shall perform the following key tasks:

[0095] Responsible for managing shared energy storage facilities within the cluster (e.g., a large battery energy storage system). ) Perform charge and discharge optimization. This involves complex energy storage state constraints:

[0096] This means that the value at the next time step (t+1) depends on the current state of charge, charging / discharging power, and corresponding efficiency. Indicates charging efficiency. This indicates a need for increased efficiency. Indicates the time step. Indicates the total capacity of the energy storage system Decide when to charge (e.g., utilizing local surplus photovoltaic power) and when to discharge it. (to meet) (Constraints or peak load within the cluster).

[0097]

[0098] Receiving area command (RM) issued After achieving the objective, coordinate the individual buildings within the cluster. Behavior. *Action assignment: Based on the energy consumption benchmarks and comfort constraints of each building Targeted load reduction or transfer instructions are issued to specific buildings within the cluster. Energy balance constraint: During optimization, energy within the cluster must be balanced.

[0099] Represents a single building At any moment The energy generated Represents a single building At any moment The generated discharge power, Represents a single building At any moment Actual energy load demand Represents a single building At any moment The charging power consumed is used for energy storage.

[0100] Event-driven, dynamically adaptive layering enables dynamic adjustment and rapid response in a two-tier architecture, with the regional command layer utilizing... Continuously monitor multi-dimensional contextual factors (real-time data, historical trends, and predicted events), and dynamically adjust the management of individual buildings based on uncertainties or triggering events. The control granularity of the model layer.

[0101] If the cluster has distributed energy resources (DERs) and microgrid functions, Energy exchange can be optimized between buildings within the cluster based on real-time supply and demand and preset rules. For example, office buildings with excess photovoltaic power can supply their surplus electricity to shared energy storage systems or laboratory buildings with high demand.

[0102] By delegating the fine-grained optimization task to the cluster-level artificial intelligence model, this method significantly improves system performance.

[0103] It is responsible for handling fine-grained data and decision-making within its jurisdiction. The Regional Command (RM) layer does not need to handle the raw data and complex interactions of all individual buildings.

[0104] Decentralized computational load: Distributing the computational burden across sub-regions .

[0105] Data isolation: Only aggregated performance metrics and capability parameters (e.g., total energy consumption within the cluster, renewable energy utilization, demand response capability, etc.) are reported back to the regional layer, rather than the raw, sensitive data. This significantly reduces the network bandwidth requirements and computational latency associated with large-scale raw data transmission and processing with the central entity.

[0106] This dynamic adjustment enables the system to respond more effectively to local events.

[0107] Rapid response: Focusing on sub-region optimization enables real-time energy arbitration, making decisions faster than global optimization models.

[0108] Enhanced resilience: In the event of localized congestion or failure, cluster-level optimization can isolate the fault, maintain power supply to critical loads within the cluster, and minimize cascading failures.

[0109] After the local optimization is completed, the aggregated results will be fed back to the regional command level through the following mechanism: Performance report: Report to the regional command (RM) on its performance in meeting requirements. Actual performance in terms of constraints and local optimization objectives.

[0110] Global Model Refinement: The Regional Command Layer (RM) utilizes these aggregated feedbacks to continuously refine its global energy optimization model through a federated learning mechanism. .

[0111] Strategy iteration: The regional command level (RM) then generates more accurate dynamic electricity price signals. Or more refined instruction This forms a highly responsive closed-loop feedback system.

[0112] The following are some different scenarios with specific descriptions. Scenario 1: Normal operation, the system is in a low-risk situation. In this scenario, the regional energy system is in a stable and predictable state, and the optimization objectives are mainly focused on minimizing operating costs and maximizing energy efficiency.

[0113] Triggering conditions and risk assessment scenarios, perception: real-time data and forecast data All results are normal.

[0114] Anomaly detection: Used by the Regional Command (RM) Model calculation status Reconstruction error .

[0115]

[0116] Risk identified: Below the preset threshold , Robustness risk level of iterative optimization evaluation for .

[0117]

[0118] Dynamic control granularity and decision execution, time-domain granularity adjustment: due to risk And error change Minimal, area command (RM) employs a relatively long control interval. (e.g., base interval) or ).

[0119]

[0120] This effectively avoids unnecessary high-frequency computational overhead.

[0121] Strategic decision-making: strategic agency ( Based on the macro-level conditions, select the normal topology configuration that offers the lowest cost and highest efficiency. .

[0122] Tactical execution: tactical agent ( Optimize fine-grained actions The regional command (RM) sends refined dynamic electricity price signals to the AI ​​models of individual buildings. .

[0123] Action space: (Continuous action vector).

[0124] Control objective: The AI ​​model of a single building is based on Adjusting energy storage power Charge and discharge rates and multi-energy assets (such as thermal storage) To minimize local operating costs.

[0125] Entropy regularization: By maximizing policy entropy This ensures that sufficient exploration efficiency can be maintained even in low-risk scenarios, thereby discovering better energy-saving strategies.

[0126] Scenario 2: Power Grid Emergency. In this scenario, the system is in a high-risk situation, facing a global, highly uncertain emergency, such as a wide-area emergency (e.g., power grid frequency). (Severe drop or large-scale failure). The system objective immediately switches to maximizing resilience and grid stability.

[0127] Triggering event: The regional command level detected real-time operational data. Power grid stability indicators (such as power grid frequency) (Exceeds the preset threshold.)

[0128] Anomaly detection: Reconstruction error A sharp rise, for example .

[0129] Risk identified: Robustness risk level of assessment Reaching the highest level, for example .

[0130] In the dynamic control granularity and decision execution process, the temporal granularity is refined: due to risks... And error change In a drastic situation, the regional command (RM) immediately implements continuous refinement of the temporal granularity. Control interval. Significantly shortened, for example from shortened to or This is to meet the sub-minute response requirements for power grid anomalies.

[0131]

[0132] in Indicates the reference control interval. Indicates the risk sensitivity coefficient. This represents the sensitivity coefficient to changes in error.

[0133] Strategic decision-making: strategic agency ( Select the topology configuration for maximum power regulation. This configuration may include enabling backup power and maximizing demand response capabilities.

[0134] Reward Shaping: reward function Weights for grid stability tilt.

[0135]

[0136] At this point, the cost coefficient Weight reduction, stability weight Significantly improved.

[0137] Tactical execution (global component-level control): Tactical agent ( Perform urgent global, component-level arbitration.

[0138] Action: Output continuous action vectors All power commands within the region. For example, requiring all critical energy storage systems in the region to... Immediately discharge at the maximum safe rate to inject stable power into the grid.

[0139] Multi-energy synergy: Adjusting heat recovery heat pumps Operating modes for multiple energy assets, in order to comply with thermoelectric coupling constraints Under the premise of achieving instantaneous load transfer.

[0140] Resilience Strategy: The Regional Command (RM) relies on insights provided by the Digital Twin Network (DTN) to predict the energy self-sufficiency of critical buildings in emergency situations and prioritize the allocation of local energy resources to these buildings.

[0141] Scenario 3: Localized load congestion and cluster interaction (localized high-risk scenario). In this scenario, RM predicts specific complex sub-regions (e.g., A localized energy supply and demand anomaly is about to occur on a university campus with shared energy assets, but the overall system remains stable.

[0142] Triggering event: The area command (RM) detected predicted event data. show Load forecasting A dramatic increase, coupled with external factors Predicted localized high temperatures will lead to increased air conditioning load. Increase.

[0143] Risk identified: error rise, Risks of assessment Reaching the intermediate level .

[0144] During the dynamic control granularity adjustment and cluster authorization process, the regional command layer initiates strategic proxy (PPO) to make macro-level decisions in order to determine the optimal topology configuration. And achieve spatial granularity refinement.

[0145] Strategic agency assessment extended status .if Predicting high temperatures in certain areas, strategic agents will utilize weather perception bias. To guide them in selecting topology configuration .

[0146]

[0147] in Based on weather indicators and bias vector It was calculated. They tend to choose topology configurations that can effectively utilize cooling assets.

[0148] Strategic Agent Selection Configuration This configuration corresponds to a temporary license. Intermediate cluster-level AI model ( ).

[0149] Activation conditions: It is activated only when there is significant internal interaction and shared energy assets in the sub-region.

[0150] Spatial granularity: The authorization mechanism enables spatial granularity refinement, dynamically transferring control and optimization focus to the user. Within the boundaries.

[0151] Tactical execution: Cluster-level fine-grained arbitration authorized cluster-level The model is responsible for conducting more detailed, real-time energy arbitration within its jurisdiction to meet the macro-level objectives of the regional command.

[0152] Cluster-level model Receive aggregated objectives issued by the regional command level The goal is to Total load limit requirement.

[0153] The regional target matching constraint must be satisfied:

[0154] Perform the following fine-grained optimization tasks within its jurisdiction: Optimize shared energy storage systems: Shared energy storage facilities are optimized for charging and discharging to balance local supply and demand and respond to regional directives. Energy storage status must comply with... Dynamic constraints.

[0155] Localized demand response coordination: Based on regional layer instructions and local events, coordinate the demand response behavior of each building within the cluster to achieve load reduction or transfer.

[0156] Peer-to-peer energy trading: Optimizing energy trading between buildings within a cluster based on real-time supply and demand and preset rules.

[0157] Multi-energy coupling arbitration: if It owns multiple energy assets, including combined heat and power (CHP) and heat recovery heat pumps (TRHP). In compliance with thermoelectric coupling constraints (e.g.) Under the premise of ), through continuous action vectors Conduct optimal arbitration.

[0158] By activating cluster-level [activations] in local high-risk scenarios The model and system achieve the following key advantages: Avoid global computational overhead: It is responsible for handling fine-grained data and decisions within its jurisdiction. This effectively avoids global computational overhead because it distributes the computational load and prevents the RM from processing raw data and complex interactions for all individual buildings.

[0159] Enhanced resilience and rapid response: By delegating refined optimization tasks, the system significantly enhances its resilience and rapid response capabilities to local emergencies. Focusing on sub-region optimization, it enables real-time energy arbitration.

[0160] Closed-loop feedback: Aggregated performance metrics and capability parameters (rather than raw sensitive data) are reported to the regional command level. The regional command (RM) uses this feedback to refine its global energy optimization model. And generate more accurate strategic regional optimization outputs (e.g. or This forms a highly responsive and intelligent closed-loop feedback system.

[0161] In all scenarios, the effectiveness of the control strategy is continuously verified and optimized through a closed-loop feedback mechanism.

[0162] Performance feedback: AI model of individual buildings and After the action is executed, aggregate the performance metrics. (For example, comfort) (As feedback)

[0163] Global model refinement: Regional command levels utilize these aggregated feedbacks through a federated averaging algorithm. Refining the global energy optimization model .

[0164] Adaptive adjustment: Refined This then generates more accurate strategic area optimization outputs (e.g., updated...). This is then distributed back to the execution layer, thus forming a highly responsive and intelligent closed-loop feedback system.

[0165] Each individual building model (client) Deployed locally on the building, it is responsible for local optimization and real-time data processing. For local data: it is responsible for collecting and processing its proprietary local real-time operational data. These data are sensitive, such as indoor temperature, occupancy rate, etc. System performance, local renewable energy generation, etc. During the local training phase, specific energy optimization sub-models are trained independently, such as load forecasting models. Control model, optimal battery scheduling model. Each individual building. Model execution: Receive instructions from the regional layer, incorporate them into the local optimization problem, and recalculate and adjust the local energy management strategy.

[0166] Collaborative intelligent sharing achieves two-layer collaboration through federated learning without accessing sensitive raw data. To address the privacy, security, and computational overhead issues associated with direct data transmission, this architecture explicitly incorporates a federated learning mechanism to achieve monolithic architecture. Privacy-preserving collaborative intelligent sharing between the model and regional models. The federated learning mechanism enhances system scalability and decentralization in the following ways: The training task of the local energy optimization sub-model is distributed to the artificial intelligence models of each individual building, thereby decentralizing the computing load. The regional command layer only aggregates model updates, avoiding the processing of large-scale raw data and reducing network bandwidth requirements and computational latency; The global energy optimization model improves the accuracy and robustness of the global model by integrating heterogeneous local learning parameters.

[0167] During local model training, the client Using local private datasets Training a local model Minimize the local loss function :

[0168] Local training is typically performed using stochastic gradient descent ( Iterate:

[0169] After the client-side training is complete, only the anonymized and encrypted model will be updated. (or gradient) Send the data to the regional aggregation server without transmitting the original sensitive data. .

[0170] Regional command levels receive and use federal averages ( The algorithm aggregates model parameters and refines the global energy optimization model. :

[0171] Refined global model Indicates the summation index, adaptively adjusting the region to optimize output. and Indicates the client index. This indicates the total number of clients and other strategic parameters, thereby forming a highly responsive and intelligent closed-loop feedback system.

[0172] standalone building Model receiving area instructions Subsequently, local energy optimization will be implemented with the goal of minimizing operating costs and carbon emissions while maximizing comfort.

[0173] in, This represents the minimization objective function, which is the comprehensive objective function that the individual building AI model seeks to minimize during local optimization. This indicates the optimization timeframe, which is the time span covered by the optimization strategy (e.g., the next 24 hours). Represents a time index, optimizing a specific time step within a time range T; This represents the operating cost weighting coefficient, used to adjust the priority of operating costs in the overall optimization objective. This represents the carbon emission weighting coefficient, used to adjust the priority of carbon emissions in the overall optimization objective; This represents the comfort weighting coefficient, used to adjust the priority of user comfort in the overall optimization objective. A negative sign before the comfort objective indicates maximizing comfort. Indicates time t The operating costs, i.e., the building's operating costs over time. t Total energy operating costs; carbon emissions for:

[0174] Indicates time t The carbon emissions of the building over time. t Total energy-related carbon emissions; Indicates time t Comfort metrics, such as penalties for indoor temperature deviating from target values; This represents the carbon intensity target (kgCO2 / kWh) refined and issued by the global model. Architecture in Time t Electricity purchased from the power grid, This represents the operating cost of the building's local power generation at time t; Energy storage system state constraints:

[0175] Among them, energy storage systems in time t State of charge The battery must always be kept within the preset minimum (SOCmin) and maximum (SOCmax) limits to prevent overcharging and over-discharging.

[0176] This means that the value at the next time step (t+1) depends on the current state of charge, charging / discharging power, and corresponding efficiency. Indicates charging efficiency. This indicates a need for increased efficiency. Indicates the time step. Indicates the total capacity of the energy storage system Event-driven, dynamically adaptive layering enables dynamic adjustment and rapid response in a two-tier architecture, with the regional command layer utilizing... Continuously monitor multi-dimensional contextual factors (real-time data, historical trends, and predicted events), and dynamically adjust the management of individual buildings based on uncertainties or triggering events. The control granularity of the model layer.

[0177] The regional command layer utilizes distributed reinforcement learning or anomaly detection algorithms to determine the most suitable spatial and temporal optimization scale based on multi-dimensional contextual factors. These multi-dimensional contextual factors include: real-time operational data, historical trend data, and predictive event data, such as future load forecasts, renewable energy generation forecasts, and potential equipment failure forecasts. Furthermore, the triggering events or predictive conditions are identified based on a comprehensive analysis of the multi-dimensional contextual factors, and include at least one of the following: grid anomalies or instability, local energy supply and demand anomalies, extreme weather warnings, specific demand response command triggers, or security threats.

[0178] Cluster level Model Aggregation target of receiving region layer And adopt model predictive control ( To conduct fine-grained energy arbitration (e.g., peer-to-peer energy trading, optimization of shared energy storage systems).

[0179] Cluster Optimize the objective function:

[0180] Energy balance constraints:

[0181] Represents a single building At any moment The energy generated Represents a single building At any moment The generated discharge power, Represents a single building At any moment Actual energy load demand Represents a single building At any moment The charging power consumed is used for energy storage.

[0182] Event-driven, dynamically adaptive layering enables dynamic adjustment and rapid response in a two-tier architecture, with the regional command layer utilizing... Continuously monitor multi-dimensional contextual factors (real-time data, historical trends, and predicted events), and dynamically adjust the management of individual buildings based on uncertainties or triggering events. The control granularity of the model layer.

[0183] Regional target matching constraints:

[0184] This mechanism, by delegating fine-grained optimization tasks, avoids global computational overhead and enhances the system's resilience and rapid response capabilities.

[0185]

[0186] standalone building The model not only performs local optimizations, but also serves as a key component in continuously updating its corresponding digital twin of a single building. ), forming a digital twin network ( Regional command level through Aggregate advanced predictive insights to enable strategic asset management.

[0187] standalone building Model Output (control action) and predictive insights Continue entering the corresponding... To calibrate and update the virtual representation: Digital twin of a single building At time step The state update formula is as follows:

[0188] in, This represents the state of the i-th individual building's digital twin at the next moment. This represents the current state of the digital twin of the i-th individual building. This represents the real-time sensor data for the i-th individual building. The i-th individual building represents the equipment operating parameters. Energy consumption optimization strategy for the i-th individual building Predictive Insights Including component performance degradation Remaining service life ( ) and potential failure conditions .

[0189] The remaining useful life (RUL) prediction model can employ a long short-term memory network (LSM). ):

[0190] Regional command layer based on aggregation Gain insights and execute strategic regional asset management, including regional predictive maintenance optimization and resilience enhancement planning.

[0191] 1. Regional predictive maintenance optimization objectives:

[0192] Limited by:

[0193]

[0194] 2. Resilience Planning Optimization Objective: The objective is to maximize the resilience index. At the same time, minimize costs and carbon emissions :

[0195] Resilience Index Measuring under various disturbance scenarios Recovery energy under critical load:

[0196] Digital Twin Network (DTN) and Strategic Intelligence • BDT Closed-Loop Updates: Each individual building's AI model continuously calibrates and updates its output. The predictive insights include component performance degradation and remaining useful life (RUL).

[0197]

[0198] • Cross-level strategy validation: RM utilizes DTN to perform multi-dimensional scenario simulations, evaluating the comprehensive impact of different regional strategies on the performance of individual buildings (e.g., energy consumption, comfort, equipment lifespan degradation). Simulation results are used to calculate key performance indicators, such as comfort. And based on the feedback, we will continue to adjust our regional strategies.

[0199]

[0200] in, Represents a single building Key performance indicators for comfort Represents a single building At any moment Indoor temperature, Represents a single building At any moment The set target temperature, This indicates the maximum permissible temperature deviation range that is acceptable to the user. II. Enhanced Dynamic Adaptive Control Implementation (ADP / DRL Fusion Drive Granularity Adjustment) RM Continuous Monitoring Extended System State Vector 1. Robust Risk Assessment and Temporal Granularity Adjustment • Risk Assessment: RM uses an autoencoder (AE) to calculate the reconstruction error E, and then performs model-free iterative optimization through adaptive dynamic programming (ADP) to generate a robust risk assessment R (e.g., R∈0,1,2). R is based on the judgment of minimizing long-term cumulative cost and no longer depends on static thresholds.

[0201] The present invention also discloses a regional energy management method based on collaborative intelligence and dynamic adaptation, characterized in that the method operates in a two-level optimization architecture, which includes a regional command layer and multiple individual building model layers. Privacy-preserving collaborative intelligent sharing is achieved through a federated learning mechanism. The federated learning includes: the multiple individual building model layers independently training local energy optimization sub-models using their local real-time operating data; the multiple individual building models generating anonymized and encrypted model updates based on the local energy optimization sub-models without transmitting the local real-time operating data; and the regional command layer receiving and aggregating the model updates to refine a global energy optimization model, and adaptively adjusting the regional optimization strategy using the global energy optimization model. By continuously calibrating and updating the digital twin network through the output of the individual building model layer, a dynamic virtual representation of the building's energy and behavior can be achieved, enabling proactive predictive management of the building. The regional command layer dynamically adjusts the control granularity of the individual building model layer or temporarily authorizes the intermediate cluster-level model based on real-time detected trigger events or predicted conditions. The intermediate cluster-level model is located between the regional command layer and the individual building model layer.

[0202] It should be understood that the processor in the embodiments of the present invention may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor described above can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0203] It is understood that the memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0204] It should be understood that the above-described memory is exemplary but not restrictive. For example, the memory in the embodiments of this application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM), etc. That is to say, the memory in the embodiments of this application is intended to include, but is not limited to, these and any other suitable types of memory.

[0205] This application also provides a computer-readable storage medium for storing computer programs.

[0206] Optionally, the computer-readable storage medium can be applied to the terminal device in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0207] This application also provides a computer program product, including computer program instructions.

[0208] Optionally, the computer program product can be applied to the terminal device in the embodiments of this application, and the computer program instructions cause the computer to execute the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of this application. For the sake of brevity, they will not be described in detail here.

[0209] This application also provides a computer program.

[0210] Optionally, the computer program can be applied to the vehicle autonomous driving device in the embodiments of this application. When the computer program is run on a computer, it causes the computer to execute the corresponding processes implemented by the terminal device in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0211] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0212] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0213] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0214] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A regional energy management system based on collaborative intelligence and dynamic adaptation, characterized in that, Specifically, it includes: It operates in a two-level optimized architecture, which includes a regional command layer and multiple individual building model layers; Privacy-preserving collaborative intelligent sharing is achieved through a federated learning mechanism. The federated learning includes: the multiple individual building model layers independently training local energy optimization sub-models using their local real-time operating data; the multiple individual building models generating anonymized and encrypted model updates based on the local energy optimization sub-models without transmitting the local real-time operating data; and the regional command layer receiving and aggregating the model updates to refine a global energy optimization model, and adaptively adjusting the regional optimization strategy using the global energy optimization model. By continuously calibrating and updating the digital twin network through the output of the individual building model layer, a dynamic virtual representation of the building's energy and behavior can be achieved, enabling proactive predictive management of the building. The regional command layer dynamically adjusts the control granularity of the individual building model layer or temporarily authorizes the intermediate cluster-level model based on real-time detected trigger events or predicted conditions. The intermediate cluster-level model is located between the regional command layer and the individual building model layer.

2. The system according to claim 1, characterized in that, The control granularity of the individual building model layer includes time granularity, spatial granularity, or operational granularity; The dynamic adjustment control granularity is achieved through an enhanced dynamic adaptive optimization mechanism, which includes: The regional command layer uses an autoencoder to calculate the system state vector. Reconstruction error ; The reconstruction error The adaptive dynamic programming model, used as the stage cost input, undergoes model-free iterative optimization to generate a robust risk assessment. With optimal risk control ; based on Change in reconstruction error Adjust the control interval continuously according to the following formula. Achieve time-domain granularity refinement and sub-minute response: in Indicates the reference control interval. Indicates the risk sensitivity coefficient. This represents the sensitivity coefficient to changes in error.

3. The system according to claim 1, characterized in that, The dynamic adjustment control granularity is further refined through a hierarchical deep reinforcement learning agent to achieve spatial granularity refinement and operational granularity adjustment, specifically including: The regional command level employs a layered strategic agent, biased based on contextual and weather perception. Select the optimal topology configuration To determine the spatial optimization scale ; Enabling hierarchical tactical agents to optimize continuous action vectors under multi-energy coupling constraints. The vector includes at least the energy storage power. Thermal storage power or cold storage power ; The tactical agent maximizes policy entropy. Ensure exploration efficiency and strategy robustness to achieve component-level fine-grained control over specific energy assets.

4. The system according to claim 1, characterized in that, The cluster-level model of the temporary authorization intermediate includes: When there are significant internal interactions and shared energy assets in a sub-region, the regional command layer intelligently activates or temporarily authorizes the intermediate cluster-level model. The intermediate cluster-level model receives the aggregation targets issued by the regional command layer. And within their jurisdiction, they shall perform localized demand response coordination or shared energy storage optimization; By delegating fine-grained optimization tasks to the intermediate cluster-level model, global computational overhead is avoided, significantly improving system resilience and rapid response capabilities.

5. The system according to claim 1, characterized in that, The federated learning mechanism enhances system scalability and decentralization in the following ways: The training task of the local energy optimization sub-model is distributed to the artificial intelligence models of each individual building, thereby decentralizing the computing load. The regional command layer only aggregates model updates, avoiding the processing of large-scale raw data and reducing network bandwidth requirements and computational latency; The global energy optimization model improves the accuracy and robustness of the global model by integrating heterogeneous local learning parameters.

6. The system according to claim 1, characterized in that, The output of the individual building model layer is continuously input to the corresponding digital twin of the individual building to calibrate and update the virtual representation. The output includes: The energy consumption optimization strategy and equipment control actions generated from the single building model; and The equipment health and life prediction generated by the single building model includes component performance degradation, remaining useful life, and potential failure conditions.

7. The system according to claim 4, characterized in that, The enhanced resilience planning and disaster response include: Using the aforementioned digital twin network, combined with external environmental data, the response of the regional power grid to various disturbances is simulated; Assess the vulnerability of regional energy infrastructure and develop robust resilience strategies, including forecasting the energy self-sufficiency of critical buildings in emergency situations.

8. The system according to claim 1, characterized in that, The digital twin network is further configured to perform cross-level scenario simulation and policy verification, including: The digital twin network is used to conduct multi-dimensional scenario simulations to evaluate the comprehensive impact of different regional strategies on the performance of individual buildings; Based on the simulation results, key performance indicators for individual buildings are calculated, including energy consumption, comfort, equipment lifespan loss, and / or carbon emissions; and The aggregated results of the aforementioned key performance indicators are used as feedback to continuously adjust and optimize regional strategies.

9. The system according to claim 1, characterized in that, The regional command layer utilizes distributed reinforcement learning or anomaly detection algorithms to determine the most suitable spatial and temporal optimization scale based on multi-dimensional contextual factors. These multi-dimensional contextual factors include: real-time operational data, historical trend data, and predictive event data, such as future load forecasts, renewable energy generation forecasts, and potential equipment failure forecasts. Furthermore, the triggering events or predictive conditions are identified based on a comprehensive analysis of the multi-dimensional contextual factors, and include at least one of the following: grid anomalies or instability, local energy supply and demand anomalies, extreme weather warnings, specific demand response command triggers, or security threats.

10. A regional energy management method based on collaborative intelligence and dynamic adaptation, characterized in that, The method operates within a two-level optimization architecture, which includes a regional command layer and multiple individual building model layers. Privacy-preserving collaborative intelligent sharing is achieved through a federated learning mechanism. The federated learning includes: the multiple individual building model layers independently training local energy optimization sub-models using their local real-time operating data; the multiple individual building models generating anonymized and encrypted model updates based on the local energy optimization sub-models without transmitting the local real-time operating data; and the regional command layer receiving and aggregating the model updates to refine a global energy optimization model, and adaptively adjusting the regional optimization strategy using the global energy optimization model. By continuously calibrating and updating the digital twin network through the output of the individual building model layer, a dynamic virtual representation of the building's energy and behavior can be achieved, enabling proactive predictive management of the building. The regional command layer dynamically adjusts the control granularity of the individual building model layer or temporarily authorizes the intermediate cluster-level model based on real-time detected trigger events or predicted conditions. The intermediate cluster-level model is located between the regional command layer and the individual building model layer.

Images