A power management system in an embedded network terminal and a control method thereof

Abstract

The application discloses a power management system in an embedded network terminal and a control method thereof, and relates to the technical field of micro-grid. The system is deployed in a building micro-grid, and can communicate with a micro-grid center control unit, a neighboring terminal or an edge server. A processing unit receives a day-ahead operation trajectory as an initial solution of rolling optimization, identifies building thermal model parameters based on local observation data, linearizes to obtain a local optimization model, fuses local short-term prediction and edge server correction information, constructs an optimization sub-problem, solves the sub-problem in a distributed manner through an alternating direction multiplier method, obtains a candidate solution of continuous variables within a calculation quota, queries a strategy mapping table for binary switch decision, and outputs the table or a limited depth search result when the solution does not converge. When communication is interrupted or the solution fails, a fallback control is executed to guarantee indoor comfort and energy storage safety. The application is used to improve scheduling real-time performance and communication efficiency under limited calculation and bandwidth conditions, thereby guaranteeing indoor comfort and energy storage safety.

Description

Technical Field

[0001] This invention relates to the field of microgrid technology, and in particular to a power management system and its control method in an embedded network terminal. Background Technology

[0002] The large-scale deployment of distributed photovoltaic (PV) and energy storage systems in office building complexes, commercial microgrids, and residential communities has driven the development of source-storage-load coordinated dispatch technology. Current mainstream solutions employ a centralized MILP / MPC (MILP mixed-integer linear programming / MPC model predictive control) framework to construct optimization models. This involves performing global mixed-integer optimization (covering air conditioning start-stop constraints, BESS state of charge constraints, power balance constraints, etc.) through a cloud / central node and issuing control commands. However, in office building scenarios, this technical approach faces multiple practical constraints:

[0003] High temperatures and humidity, coupled with rapid cloud cover changes in South China, lead to significant fluctuations in photovoltaic output and building air conditioning loads (especially in office areas and residential communities, depending on usage time and load intensity). Furthermore, the thermal parameters (such as heat capacity and equivalent thermal resistance) of residential and office buildings within a building complex exhibit strong regional variations due to structural differences, making it easy for fixed model parameters to cause prediction biases. Embedded terminals (such as building edge controllers) are limited by processor performance and memory capacity, making it difficult to solve mixed-integer problems involving coupling constraints across multiple building types in real time. When communication links are abnormal (such as bandwidth fluctuations or interruptions), cloud-based decision-making reliance prevents timely responses from strategies such as office building air conditioning management and energy storage regulation, impacting thermal comfort and equipment safety in the production environment. Centralized architectures require uploading sensitive data such as residential user room temperature curves and office building equipment start-up and shutdown behaviors, posing privacy and compliance risks. Additionally, the expansion of system scale (N buildings × M types of load) leads to a non-linear increase in computational complexity and communication overhead, resulting in decreased real-time performance. During communication interruptions, residential loads may violate comfort constraints, while critical areas of office buildings (such as data center server rooms and meeting spaces) are prone to temperature control instability and other safety hazards due to the lack of adaptive operating strategies.

[0004] While existing technologies have improved scheduling economy through thermal resistance-thermal capacity equivalent modeling and time-series prediction, their core assumptions focus on the homogeneous load of residential communities, and are insufficiently adaptable to complex operating conditions such as the nonlinear thermal response of office buildings and multi-period energy consumption patterns. Traditional centralized / cloud-centred architectures face the risk of increased failure rates under the constraints of embedded terminal computing power and poor regional communication stability. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention provides a power management system and its control method in an embedded network terminal. The aim is to provide a distributed rolling optimization power management scheme for embedded network terminals in building microgrids, integrating online identification, predictive correction fusion, compressed communication, and integer decision backoff mechanisms. This scheme aims to improve scheduling real-time performance and communication efficiency, enhance privacy protection and fault robustness under limited computing and bandwidth conditions, thereby ensuring indoor comfort and energy storage safety.

[0006] The specific technical solution is as follows:

[0007] An embedded network terminal power management system, deployed in a building microgrid, is capable of communicating with the microgrid central control unit, neighboring terminals, or edge servers. It includes a processing unit, a storage unit, a sensor interface, an execution interface, and a communication module. The processing unit receives the day-ahead operating trajectory from the microgrid central control unit and uses it as the initial solution for rolling optimization. Based on local observation data collected through the sensor interface, it performs online identification, estimates building thermal model parameters, and linearizes them to obtain a local optimization model for the current rolling period. It generates a local short-term forecast and fuses this forecast with correction information from the edge server to form a model for rolling optimization. The system integrates prediction and calculation. Based on the initial solution, the local optimization model, and the integrated prediction, a local optimization sub-problem is constructed. Compressed aggregated information is exchanged with the microgrid central control unit or the neighboring terminal. Distributed solution is performed using the alternating direction multiplier method to obtain continuous variable candidate solutions within a limited computational quota. When the local optimization sub-problem involves binary switch decisions, an integer action is obtained by querying the strategy mapping table. If the distributed solution fails to reach the convergence criterion within the computational quota, the output of the strategy mapping table or the result of a finite depth search is used as the backoff integer decision. In the event of communication interruption or local solution failure, backoff control is executed to ensure indoor comfort and energy storage safety. Event-triggered reporting is also supported.

[0008] Furthermore, in the above scheme, the online identification uses recursive least squares or extended Kalman filtering to estimate the parameters of the second-order thermodynamic model of the dual-resistor dual-capacitor system, and updates the linearization coefficients of the next rolling period at the end of each rolling cycle.

[0009] Furthermore, in the above scheme, the local short-term prediction adopts any one of the following: long short-term memory neural network, autoregressive moving average model, or autoregressive model; the correction information fusion is performed according to the weighted average rule or Bayesian update rule, and the fusion weight is determined according to the issued systematic correction amount and prediction confidence information according to the predefined mapping.

[0010] Furthermore, in the above scheme, when using the alternating direction multiplier method for distributed solution, the exchanged information undergoes sparse transformation, quantization, and differential encoding to form a compressed Lagrange vector or intention vector; the computation quota is dynamically determined by the embedded network terminal according to predefined rules, which take processor utilization, remaining decision time limit, and network latency as inputs to generate the maximum number of iterations or total computation time limit.

[0011] Furthermore, the above scheme further includes a strategy mapping table generated by a cloud or edge server based on offline mixed-integer linear programming simulations and historical running data. This involves solving mixed-integer linear programming problems across multiple scenarios to obtain a sample set of features and actions. An initial mapping model is then trained using feature selection and supervised learning, followed by pruning, quantization, or knowledge distillation to create a decision tree, neural network, or lookup table, which is then distributed to an embedded network terminal. When the output of the strategy mapping table does not meet the allowable range of indoor temperature or the upper and lower limits of the energy storage state of charge, a finite-depth search is performed to correct the integer actions until the constraints are met or the search limit is reached. The strategy mapping table includes version information and failure rules when it is distributed.

[0012] Furthermore, when using finite-depth search as a backoff strategy, the search depth is limited by a predefined upper limit; candidate integer actions are scored based on a weighted cost function that includes switching penalties, indoor temperature deviation, and energy costs; the weights of the cost function are given by system configuration parameters; and the final action must meet the allowable range of indoor temperature and the upper and lower limits of the energy storage state of charge.

[0013] Furthermore, the above scheme includes the following conditions for triggering event reporting: the absolute error between the local short-term forecast and the day-ahead operating trajectory exceeds a predetermined threshold; the equipment operating status is abnormal; the energy storage state of charge exceeds the preset upper and lower limits; when reporting is triggered, the reported quantity is differentially encoded and compressed, which includes three steps: transformation according to a preset sparse transformation, quantization using fixed-point quantization, and encoding using a differential method; when privacy protection is required, differential privacy noise of a predetermined strength is injected; the encoded data stream is transmitted in an encrypted manner.

[0014] Furthermore, in the event of communication interruption or distributed solution failure, the fallback control adopts simplified model predictive control or rule control: the simplified model predictive control uses linearized thermal balance constraints as the model, and transforms the optimization of continuous control variables into quadratic programming to generate feasible control sequences.

[0015] Furthermore, the above scheme

[0016] The processing unit also includes a data acquisition and preprocessing module, an online parameter identification module, a short-term prediction and fusion module, a rolling optimization and distributed solution module, a strategy mapping table query module, a rollback control module, and a communication and privacy module.

[0017] Secondly, this invention discloses a control method for a power management system in an embedded network terminal.

[0018] The system receives the day-ahead operating trajectory from the microgrid central control unit and uses the day-ahead operating trajectory as the initial solution for rolling optimization; it performs online identification based on local observation data, estimates building thermal model parameters, and linearizes and constructs a local optimization model for the current rolling period accordingly.

[0019] A local short-term forecast is generated and fused with correction information from the edge server to form a fused forecast for rolling optimization.

[0020] Based on the initial solution, the local optimization model, and the fusion prediction, a local optimization subproblem is constructed. Compressed aggregated information is exchanged with the microgrid central control unit or a nearby terminal. The alternating direction multiplier method is used for distributed solution to obtain continuous variable candidate solutions within the limited computational quota.

[0021] When the local optimization subproblem involves binary switch decisions, the policy mapping table is queried to obtain integer actions. If the distributed solution fails to reach the convergence criterion within the computational quota, the output of the policy mapping table or the result of the finite depth search is used as the backoff integer decision.

[0022] In the event of communication interruption or local solution failure, rollback control is executed to ensure indoor comfort and energy storage safety; aggregated quantity or compressed Lagrange information is reported in an event-triggered manner, and differential coding and compression coding are used to reduce communication volume.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This invention significantly improves the adaptability of the local model to the non-ideal characteristics of actual buildings and equipment by identifying and updating the linearized model online in each rolling cycle, ensuring the solvability of rolling optimization under embedded computing resources. By fusing local prediction with cloud / edge correction information, the impact of prediction errors on scheduling decisions is reduced, improving energy utilization efficiency and comfort maintenance capabilities. By introducing compression strategies such as sparse transformation, quantization, and differential coding in distributed solving, combined with event-triggered reporting, communication bandwidth requirements and reporting frequency are significantly reduced. At the same time, the fine-grained temperature timing or device behavior sequence of a single terminal is avoided in information exchange, thereby improving privacy protection capabilities. The combination of policy mapping table and finite depth search provides an efficient backoff path for integer switching decisions within the limited computing quota, enhancing the stability and real-time response capability of the solution. The backoff control design ensures operational robustness in the event of communication interruption or solution failure, maintaining indoor comfort and energy storage safety. Overall, it balances real-time performance, communication efficiency, privacy protection, and system robustness.

[0025] The above mechanisms work together in the modular software implementation of the embedded terminal: including data acquisition and preprocessing, online parameter identification, short-term prediction and correction fusion, rolling optimization and distributed solution, strategy mapping table query, rollback control, communication and privacy modules, etc., so as to realize reliable, scalable and user privacy-protecting distributed power and heat load management in building microgrid scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the functional modules of the present invention;

[0027] Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0028] The embodiments of the invention will be described in further detail below with reference to the accompanying drawings, so that the objectives, technical solutions and technical effects of the invention will be more clearly presented.

[0029] like Figures 1-2As shown, this invention discloses an embedded network terminal power management system deployed in a building microgrid. It can communicate with the microgrid central control unit, neighboring terminals, or edge servers. The system includes a processing unit, a storage unit, a sensor interface, an execution interface, and a communication module. The processing unit is configured to execute a program stored in the storage unit to: receive the day-ahead operating trajectory from the microgrid central control unit and use it as the initial solution for rolling optimization; perform online identification based on local observation data collected through the sensor interface, estimate building thermal model parameters, and linearize them to obtain a local optimization model for the current rolling period; generate a local short-term forecast and compare the forecast with the calibration data from the edge server. Positive information fusion is used to form a fused prediction for rolling optimization. Based on the initial solution, local optimization model, and fused prediction, a local optimization subproblem is constructed. Compressed aggregated information is exchanged with the microgrid central control unit or nearby terminals. Distributed solution is performed using the alternating direction multiplier method to obtain continuous variable candidate solutions within the limited computational quota. When the local optimization subproblem involves binary switching decisions, the strategy mapping table is queried to obtain integer actions. If the distributed solution fails to reach the convergence criterion within the computational quota, the output of the strategy mapping table or the result of the finite depth search is used as the backoff integer decision. In the event of communication interruption or local solution failure, backoff control is executed to ensure indoor comfort and energy storage safety. Event-triggered reporting is also supported.

[0030] The working principle of this invention will be explained below:

[0031] First, the embedded network terminal performs online identification based on local observation data collected by sensors, and uses recursive least squares method or extended Kalman filter to estimate the parameters of the second-order thermodynamic model of dual resistor and dual capacitor. At the end of each rolling cycle, the linearization coefficients of the next rolling period are updated accordingly to ensure that the local optimization model can reflect recent operating characteristics and is easy to solve on the embedded platform.

[0032] Secondly, the terminal uses methods such as Long Short-Term Memory Network (LSTM), Autoregressive Moving Average (ARMA) or Autoregressive (AR) model to generate local short-term predictions, and then merges these predictions with systematic correction information provided by the cloud or edge server according to weighted average or Bayesian update rules. The fusion weights are determined by the issued systematic correction terms and the confidence of the local predictions through a predefined mapping, thereby obtaining corrected predictions for rolling optimization.

[0033] Furthermore, based on the initial solution of the current day's trajectory, the updated linearized local model, and the corrected prediction, a local rolling optimization subproblem is constructed at the terminal. The alternating direction multiplier method is used to exchange compressed Lagrangian information or intent vectors processed by sparse transformation, quantization, and differential coding with the microgrid center or nearby terminals within the limited computing quota, so as to realize the solution of distributed continuous variables. At the same time, real-time performance is guaranteed by dynamically determined computing quota (with processor utilization, remaining decision time limit, and network latency as inputs).

[0034] For integer decisions involving binary switches, this invention pre-configures a strategy mapping table on the terminal side, which is generated offline and distributed by a cloud or edge server. The mapping table is trained using multi-scenario mixed integer linear programming simulation and historical data, and then processed through pruning, quantization, or knowledge distillation to become a decision tree, lightweight neural network, or lookup table structure. When the distributed solution fails to meet the convergence criterion within the computational quota or the mapping table output does not meet the temperature and state of charge constraints, the terminal can perform a finite-depth search at a predefined depth to correct the integer action. The search evaluates candidate actions using a weighted cost function that includes switching penalties, temperature deviations, and energy costs, and selects a feasible solution that meets the constraints.

[0035] To address communication interruptions or local solution failures, this invention designs a backoff control strategy: a simplified linearized thermal balance constraint model predictive control or rule-based control is used to transform continuous variable optimization into quadratic programming to generate feasible continuous control sequences, while integer actions are still determined by a strategy mapping table or finite-depth search to ensure indoor comfort and energy storage safety.

[0036] The building thermal model structure adopts a second-order thermodynamic model with two resistors and two capacitors (2R–2C). Parameters θ include: indoor air heat capacity C_a, wall equivalent heat capacity C_w, indoor-wall equivalent thermal resistance Ra_aw, and indoor-outdoor equivalent thermal resistance Rao. The state x = [T_a, T_w]^T, and the input u includes air conditioning cooling / heating power P_HVAC and external disturbances (outdoor temperature T_out, solar irradiance G). Discretized with a sampling period Δt, we obtain x_{k+1} = A(θ,Δt)x_k + B(θ,Δt)u_k + E(θ,Δt)d_k; y_k = Cx_k. After obtaining θ̂ online, at the end of each rolling cycle, first-order linearization is performed according to the current operating point to form linearization constraints for the next rolling period. Recursive least squares method (forgetting factor 0.98–0.995) or extended Kalman filter (process / measurement noise covariance can be adaptively adjusted according to the residual); input features include indoor temperature, wall surface temperature (or its approximate measurement), air conditioning power, outdoor temperature, and solar irradiance; outliers are removed by 3σ test.

[0037] This invention relates to an embedded network terminal deployed in a building microgrid. The terminal hardware includes: a processing unit, which can be an ARM or x86 embedded processor with a floating-point unit, typically with a clock speed of 0.6–1.5 GHz; a storage unit, which can be flash memory or RAM, used to store programs, policy mapping tables, and logs; a sensor interface for supporting the acquisition of multiple temperature, power, and status signals; an execution interface for outputting control signals to the drive circuits or communication ports of air conditioning, heat sources, and energy storage devices; and a communication module, which can adopt Ethernet / Wi-Fi / LoRa / RS485 / 4G, etc. The embedded network terminal's functional modules include: a data acquisition and preprocessing module, an online identification module, a short-term prediction and correction fusion module, a rolling optimization and distributed solution module, a policy mapping table query module, a finite-depth search module, a rollback control module, a communication and privacy module, and an operation log management module. The implementation details and parameter examples of each module will be explained item by item below.

[0038] Data acquisition and preprocessing module

[0039] This module is responsible for periodically collecting various observations (indoor temperature, building envelope surface temperature, outdoor temperature, solar irradiance, equipment on / off status, energy storage voltage / current / SoC, etc.) from terminal sensor interfaces, and processing the raw data into usable inputs for identification, prediction, and optimization. Specific tasks include: setting the sampling period (example: 5 min for indoor temperature, or 30 s–5 min); low-pass filtering (e.g., first-order exponential smoothing, α=0.2–0.5); outlier detection (sliding window midpoint filtering or 3σ test); aggregation by decision step size (e.g., 5 min or 15 min) and incorporating it into the local cache; generating event judgment signals to trigger reporting (e.g., prediction and day-ahead trajectory error exceeding threshold, sensor or equipment malfunction, SoC out of bounds). The module outputs: preprocessed time-series data, event trigger flags, and reporting candidate packets.

[0040] The parameters of the current operating trajectory are as follows: time series fields are arranged according to the rolling optimization step size Δt (configurable from 5 to 15 min): indoor temperature upper and lower limits reference T_in_min[k], T_in_max[k]; air conditioning power or on / off reference P_HVAC_ref[k] or u_HVAC_ref[k]; energy storage charging and discharging reference P_BESS_ref[k] and state of charge reference SoC_ref[k]; maximum demand / demand response constraints P_max[k] / DR_window[k]; electricity price sequence π[k]; external disturbance references such as outdoor temperature T_out_ref[k] and solar irradiance G_ref[k]. Each sequence is accompanied by a timestamp, version number and validity period.

[0041] Acquisition and distribution method: The microgrid center control unit generates the solution after system-level optimization based on the day-ahead market clearing results, weather forecast, and park operation constraints. It then distributes the solution to the terminal via an encrypted channel through the communication module. The terminal loads the solution after verifying the timestamp, signature, and version number, and stores it in the local cache as the initial solution segment for rolling optimization.

[0042] Online identification module: The online identification uses recursive least squares or extended Kalman filtering to estimate the parameters of the second-order thermodynamic model of the two-resistor and two-capacitor system, and updates the linearization coefficients of the next rolling period at the end of each rolling cycle.

[0043] A second-order thermodynamic model with two resistors and two capacitors (2R–2C) is adopted as an approximate model for the indoor temperature dynamics of buildings. The model can be expressed as a continuous-time state equation, which is discretized for online identification and linearization. The model formula is as follows:

[0044] x_{k+1}=A(θ)x_k+B(θ)u_k+w_k;

[0045] y_k = C x_k + v_k;

[0046] Where x is the state vector, representing indoor temperature and wall temperature, etc.; u is the control input, representing air conditioning cooling / heating power and external disturbances such as solar radiation and ambient temperature; θ is the parameter to be identified, such as heat capacity and thermal resistance; w_k and v_k are noise. To ensure real-time performance on the embedded platform, the model can be linearized and stored in a linear form in the state space.

[0047] Online parameter estimation can be performed using existing recursive least squares (RLS) or extended Kalman filtering (EKF), which will not be elaborated here.

[0048] Short-term forecasting and correction fusion module:

[0049] Local short-term forecasts employ any one of the following: long short-term memory neural network, autoregressive moving average model, or autoregressive model; fusion is performed according to the weighted average rule or Bayesian update rule, and the fusion weights are determined according to the issued systematic correction amount and forecast confidence information based on a predefined mapping.

[0050] Short-term predictions can employ LSTM, ARMA, or AR models. A lightweight LSTM model can be deployed when local resources allow; AR or ARMA models can be used when resources are limited. Input features include historical indoor temperature sequences, the last control action, external disturbances (outdoor temperature, solar irradiance), and time information (hours, weekdays / weekends). Training methods include offline and online fine-tuning: in the offline stage, the model's initial weights can be obtained by training on historical data in the cloud or on an edge server, and a simplified model can be distributed to the terminal; in the online stage, small-batch ADAM updates or incremental parameter adjustments can be used to adapt to changes.

[0051] The correction information comes from system-level prediction biases or edge update signals from the cloud or edge servers, representing system biases in the macro environment / group behavior, such as system errors in weather forecasts.

[0052] Fusion Method Steps: Weighted average fusion or Bayesian update. Specific Implementation: Let the local prediction be ŷ_local, and the correction information be Δ_sys (issued by the center). The fused prediction is ŷ = w_local·ŷ_local + w_sys·(ŷ_local + Δ_sys). The weight w_sys is issued by the center or determined through a predefined mapping function f(confidence_local,confidence_sys). Confidence_local can be estimated from the uncertainty inherent in the prediction model, such as the dropout uncertainty of LSTM or the residual variance of AR. Confidence_sys is provided by the center. During Bayesian update, if both the local prediction and the system correction are considered Gaussian distributions, they can be updated using the mean-variance merging formula. Example of weight range: w_sys ∈ [0, 0.5]. Increase w_sys when the local uncertainty is high.

[0053] Rolling optimization and distributed solution module

[0054] When using the alternating direction multiplier method for distributed solution, the exchanged information undergoes sparse transformation, quantization, and differential encoding to form a compressed Lagrange vector or intention vector. The computation quota is dynamically determined by the embedded network terminal according to predefined rules, which take processor utilization, remaining decision time limit, and network latency as inputs to generate the maximum number of iterations or total computation time limit.

[0055] In each rolling cycle, a local rolling optimization subproblem is constructed. The prediction window N is configurable, such as N=12. Variables include continuous quantities, such as air conditioning power and energy storage charging / discharging power, and integer quantities, such as equipment on / off. The day-ahead trajectory is used as the initial solution to accelerate convergence. Distributed solving employs ADMM splitting, with the entire network or neighboring terminals / centers alternately solving and exchanging Lagrange information. A penalty parameter ρ is chosen, ρ=0.1–10, and the multiplier λ is initialized. For local problems of QP, an embedded solver is recommended; in resource-constrained situations, projected gradient approximation can be used. The communication layer compresses the exchanged Lagrange information: DCT / wavelet sparse transform → fixed-point quantization 8–12 bits → differential coding → encryption. Quota calculation is dynamically decided locally, with the iteration upper limit determined based on CPU usage, remaining decision time T_remain, and network latency RTT. For example, the mapping is: MaxIter=20 when CPU < 40%, MaxIter=10 when 40–70%, and MaxIter=5 when > 70%. The rolling optimization and distributed solution module outputs candidate solutions for continuous variables, local consistency residual indices, and contextual information for integer decision-making.

[0056] Strategy Mapping Table Query Module

[0057] The policy mapping table is generated by a cloud or edge server based on offline mixed-integer linear programming simulations and historical running data. The steps include: obtaining a sample set of features and actions by solving mixed-integer linear programming problems across multiple scenarios; training an initial mapping model through feature selection and supervised learning; then processing the model through pruning, quantization, or knowledge distillation into a decision tree, neural network, or lookup table format, and finally distributing it to an embedded network terminal. When the output of the policy mapping table does not meet the allowable range of indoor temperature or the upper and lower limits of the energy storage state of charge, a finite-depth search is performed to correct the integer actions until the constraints are met or the search upper limit is reached. The policy mapping table includes version information and failure rules upon distribution.

[0058] When computation is limited or fast integer decision-making is required, the local policy table is invoked first. The policy table is generated offline from the cloud / edge: batch solving of multi-scenario MILP to generate feature-action samples, trained using supervised learning methods such as decision trees / small neural networks, and then pruned, quantized, or knowledge distilled and compressed into a lookup table or lightweight inference model suitable for terminal deployment. Table features include: current indoor temperature, predicted load, energy storage SoC, time period identifier, previous day trajectory action, recent handover frequency, etc. A version number and failure rules are included with the deployment; a failure flag is returned when the input exceeds the training range. At runtime, input features are used, and candidate integer actions and confidence scores are output; if an action is infeasible (violates temperature or SoC constraints), a failure is returned and a finite-depth search is triggered. The policy table query module is used to output candidate integer actions or failure flags.

[0059] Finite Depth Search Module

[0060] When using a finite-depth search as a backoff strategy, the search depth is limited by a predefined upper limit. Candidate integer actions are scored based on a weighted cost function that includes switching penalties, indoor temperature deviation, and energy costs. The weights of the cost function are given by system configuration parameters; the final action must meet the allowable indoor temperature range and the upper and lower limits of the energy storage state of charge.

[0061] When the mapping table has no actions or the distributed solution has not converged and requires integer correction, the terminal performs a finite-depth search within a predefined depth D (D=2 or 3). Using the mapping table output or the nearest feasible solution as the root, neighborhood candidates are enumerated. For each candidate, the cost is quickly evaluated and constraints are checked using a linearized model in the short term. The cost function is: cost = α·switching penalty + β·temperature deviation penalty + γ·energy cost (weights α=1.0, β=10.0, γ=0.5). Heuristic pruning and relaxation decisions are used to control the computation. If the search finds an action within the quota that satisfies the constraints and has the minimum cost, that action is returned; otherwise, failure is returned, triggering a fallback control. This module outputs: the corrected integer action or a failure flag.

[0062] The finite depth search and rollback control must meet the following constraints when filtering or generating actions; if the constraints cannot be met simultaneously, the priority is downgraded to safety > comfort > economy.

[0063] Allowable indoor temperature range: During working hours, T_in ∈ [22℃, 26℃], and during non-working hours, it can be relaxed to [18℃, 30℃]; the specific time-sharing table is stored in the terminal configuration and can be updated remotely.

[0064] Energy storage state of charge upper and lower limits: SoC∈[10%,95%], and set constraints such as charging and discharging power, ramp rate and efficiency; when the safety threshold is exceeded, the safety constraints are given priority.

[0065] rollback control module

[0066] The conditions for triggering an event report include any of the following: the absolute error between the local short-term forecast and the day-ahead operating trajectory exceeds a predetermined threshold; the equipment operating status is abnormal; the energy storage state of charge exceeds the preset upper and lower limits; when a report is triggered, the reported quantity is differentially encoded and compressed, which includes three steps: transformation according to a preset sparse transformation, quantization using fixed-point quantization, and encoding using a differential method; when privacy protection is required, differential privacy noise of a predetermined strength is injected; the encoded data stream is transmitted in an encrypted manner.

[0067] Triggered in the event of communication interruption, distributed solution failure, or both the policy mapping table and search fail, this rollback module is responsible for ensuring minimum guarantees of indoor comfort and energy storage safety. Rollback strategies include simplified linear MPC (transforming continuous variable problems into QP (quadratic programming)) or rule-based control. Simplified MPC design: The objective is to minimize the quadratic cost of temperature deviation and energy consumption, with constraints including linearized thermal balance and device power / SoC boundaries; the QP solver can be OSQP (open-source quadratic programming solver) or a lightweight built-in solver. Integer actions are preferentially provided by the policy mapping table or finite-depth search; if none are available, the existing switching state is maintained or executed according to empirical rules (such as continuing the action from the previous moment or following a minimum switching strategy). The rollback control module outputs executable continuous control sequences and combinations of integer actions until normal solution capability is restored.

[0068] Communication and Privacy Module

[0069] When communication is interrupted or distributed solution fails, the backoff control adopts simplified model predictive control or rule control: the simplified model predictive control uses linearized thermal equilibrium constraints as the model and transforms the optimization of continuous control variables into a quadratic programming solution to generate feasible control sequences.

[0070] The communication and privacy module is responsible for event-triggered reporting, compression encoding, differential privacy noise injection, encryption, and version management. Event triggering rules include: local prediction and day-ahead trajectory absolute error exceeding the threshold ε_pred≈0.5–1.0℃, device malfunction, and SoC out-of-bounds (e.g., SoC <10% or >95%). The reported data processing flow is as follows: sparse transformation (DCT / WT retains major coefficients) → fixed-point quantization (multiplying by the quantization factor and integerizing) → differential encoding (inter-frame increment) → differential privacy noise injection (if configured, Laplace mechanism: noise follows Laplace(0,Δ / ε), ε can be configured such as 0.1–1.0) → packaging and encryption (AES-128 / 256 or TLS). The message header should include the compression method, quantization factor, timestamp, and version number for the receiver to decode and verify. The module is also responsible for resending key summary logs or necessary aggregation information when communication is restored. The module outputs: encrypted and compressed reporting packets and communication status indicators.

[0071] Simplified Model Predictive Control (Simplified MPC):

[0072] The objective function is J = Σ_k(w_T·(T_a(k)-T_set(k))^2+w_P·P_HVAC(k)^2+w_S·Δu(k)^2); where w_T is the temperature deviation weighting coefficient, T_a(k) is the measured indoor temperature at time k, T_set(k) is the temperature setpoint at time k, w_P is the energy consumption weighting coefficient, P_HVAC(k) is the air conditioning power at time k, w_S is the control action smoothing weighting coefficient, and Δu(k) is the change in control quantity at time k.

[0073] The constraints include linearized thermal balance, device power upper and lower limits, SoC upper and lower limits, and ramping constraints; continuous variable optimization is transformed into QP (quadratic programming) solution; the solver can use OSQP (open source quadratic programming solver), and the maximum solution time limit and tolerance are configurable (e.g., time limit 50–150 ms, ρ adaptive).

[0074] Integer action: given by the policy mapping table. If the mapping table fails, a finite depth search is used. If it is still not feasible, the action of the previous control step is maintained or the minimum switching policy is executed.

[0075] Run log management module

[0076] This module is responsible for locally recording critical operational data for auditing, fault diagnosis, and re-reporting: parameter identification history, prediction residuals, each optimization and search result, event trigger records, reporting records, and communication disconnection records. It implements local log rotation, persistent storage, and a priority-based re-reporting mechanism, allowing selective re-reporting based on time windows or event importance after communication is restored. It also supports reading historical states during mapping tables or firmware rollbacks to restore consistency. The module's output includes a local log index, a re-reporting queue, and a status report.

[0077] The steps for the above solution can be described as follows:

[0078] Receive the current day's running trajectory and initialize the initial solution for rolling optimization; start local data acquisition and preprocessing.

[0079] Perform online identification and estimation of thermal model parameters and linearize them into a local optimization model.

[0080] The system performs short-term forecasts locally and fuses the issued system correction information with predefined fusion rules to obtain forecasts for decision-making.

[0081] A local rolling optimization subproblem is constructed. Within a specified computational quota, compressed Lagrangian information is exchanged with the central or neighboring terminals using ADMM, and candidate solutions for continuous variables are solved. At the end of each iteration, the remaining time is checked to determine whether to continue.

[0082] If the subproblem involves binary switch decisions, the local policy mapping table is queried first to obtain integer actions; if the mapping table output does not meet the constraints or the distributed solution does not converge, a finite depth search is performed within the predefined depth and actions that meet the constraints are selected using a weighted cost function.

[0083] If a feasible integer solution cannot be obtained within the calculation quota or search limit, or if communication is interrupted, backoff control is triggered to generate a feasible control sequence and ensure comfort and energy storage safety.

[0084] The compressed aggregate quantity or Lagrange information is reported in parallel according to the event triggering rules. Before reporting, it is processed according to the compression process and differential privacy noise is injected as needed. The data is sent through an encrypted channel. After communication is restored, necessary logs and information are supplemented.

[0085] like Figure 2 As shown, this invention also discloses a control method for a power management system in an embedded network terminal, the specific method steps of which include:

[0086] S1. Receive the day-ahead operating trajectory from the microgrid central control unit and use the day-ahead operating trajectory as the initial solution for rolling optimization.

[0087] S2. Based on local observation data, perform online identification, estimate building thermal model parameters, and linearize them to construct a local optimization model for the current rolling period.

[0088] S3. Generate local short-term forecasts and fuse them with correction information from the edge server to form a fused forecast for rolling optimization.

[0089] S4. Based on the initial solution, local optimization model and fusion prediction, construct a local optimization subproblem, exchange compressed aggregated information with the microgrid central control unit or nearby terminal, and use the alternating direction multiplier method for distributed solution to obtain continuous variable candidate solutions within the limited computation quota.

[0090] S5. When the local optimization subproblem involves binary switch decisions, query the policy mapping table to obtain integer actions, and if the distributed solution fails to reach the convergence criterion within the computational quota, use the output of the policy mapping table or the result of the finite depth search as the backoff integer decision.

[0091] S6. In the event of communication interruption or local solution failure, execute rollback control to ensure indoor comfort and energy storage safety; report aggregated quantity or compressed Lagrange information in an event-triggered manner, and use differential coding and compression coding to reduce communication volume.

[0092] To verify the effectiveness of this solution, a microgrid field test was conducted in an office park with multiple buildings. The results show that, compared to traditional centralized control, this system, with limited computing power in embedded terminals, achieves a convergence probability of over 95% within 5 seconds using its distributed optimization algorithm, reducing the communication load between adjacent terminals by approximately 70%. In simulated communication interruption fault tests, the system seamlessly switched to fallback control mode, successfully controlling indoor temperature fluctuations within ±0.5℃ of the set value and ensuring the safe operation of the energy storage unit. Simultaneously, through event-triggered mechanisms and data compression technology, the data traffic reported to the central node was reduced by more than 60%, significantly improving privacy protection capabilities. The experimental data fully demonstrates the comprehensive advantages of this system in terms of real-time performance, communication efficiency, robustness, and privacy security.

[0093] The English-Chinese translation of the above terminology is as follows:

[0094] Alternating Direction Multiplier Method (ADMM)

[0095] Adaptive Moment Estimation Optimization Algorithm (ADAM)

[0096] Mixed Integer Linear Programming (MILP)

[0097] Model predictive control (MPC)

[0098] Quadratic Programming (QP)

[0099] Open source quadratic programming solver (OSQP)

[0100] Long Short-Term Memory (LSTM) network

[0101] Autoregressive / Autoregressive Moving Average (AR / ARMA)

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the patent application of the present invention. All equivalent changes, substitutions or modifications made within the technical spirit and principles indicated by the present invention should be included within the scope of patent protection covered by the present invention.

Claims

1. An embedded network terminal power management system, deployed in a building microgrid, capable of communicating with the microgrid central control unit, nearby terminals, or edge servers, comprising a processing unit, a storage unit, a sensor interface, an execution interface, and a communication module; characterized in that: The processing unit receives the day-ahead operating trajectory from the microgrid central control unit and uses the day-ahead operating trajectory as the initial solution for rolling optimization; Based on the local observation data collected through the sensor interface, online identification is performed to estimate the building thermal model parameters and linearize them to obtain a local optimized model for the current rolling period. Using any one of the following: long short-term memory neural network, autoregressive moving average model, or autoregressive model, and based on input features including historical indoor temperature, control actions, external disturbances, and time information, a local short-term prediction is generated. This local short-term prediction is then fused with correction information from the edge server to form a fused prediction for rolling optimization. Based on the initial solution, the local optimization model, and the fusion prediction, a local optimization subproblem is constructed. Compressed aggregated information is exchanged with the microgrid central control unit or the neighboring terminal. Distributed solution is performed using the alternating direction multiplier method, obtaining continuous variable candidate solutions within a limited computational quota. When the local optimization subproblem involves binary switching decisions, an integer action is obtained by querying the strategy mapping table. If the distributed solution fails to reach the convergence criterion within the computational quota, the output of the strategy mapping table or the result of a finite-depth search is used as the backoff integer decision. In the event of communication interruption or local solution failure, backoff control is executed to ensure indoor comfort and energy storage safety. Event-triggered reporting is supported. The conditions for event-triggered reporting include any of the following: the absolute error between the local short-term prediction and the day-ahead operating trajectory exceeds a predetermined threshold; the equipment operating status is abnormal; the energy storage state of charge exceeds preset upper or lower limits.

2. The embedded network terminal power management system according to claim 1, characterized in that: The online identification uses recursive least squares or extended Kalman filtering to estimate the parameters of the second-order thermodynamic model of the dual-resistor dual-capacitor system, and updates the linearization coefficients for the next rolling period at the end of each rolling cycle.

3. The embedded network terminal power management system according to claim 1, characterized in that: The correction information fusion is performed according to the weighted average rule or the Bayesian update rule, and the fusion weight is determined according to the issued systematic correction amount and prediction confidence information according to the predefined mapping.

4. The embedded network terminal power management system according to claim 1, characterized in that: When using the alternating direction multiplier method for distributed solution, the exchanged information undergoes sparse transformation, quantization, and differential encoding to form a compressed Lagrange vector or intention vector. The computation quota is dynamically determined by the embedded network terminal according to predefined rules, which take processor utilization, remaining decision time limit, and network latency as inputs to generate the maximum number of iterations or total computation time limit.

5. The embedded network terminal power management system according to claim 1, characterized in that: The strategy mapping table is generated by a cloud or edge server based on offline mixed-integer linear programming simulation and historical running data: a sample set of features and actions is obtained by solving mixed-integer linear programming in multiple scenarios, an initial mapping model is obtained through feature selection and supervised learning training, and then processed into decision trees, neural networks or lookup tables through pruning, quantization or knowledge distillation, and then distributed to embedded network terminals; when the output of the strategy mapping table does not meet the allowable range of indoor temperature or the upper and lower limits of the energy storage state of charge, a finite-depth search is performed to correct the integer actions until the constraints are met or the search upper limit is reached; the strategy mapping table includes version information and failure rules when it is distributed.

6. The embedded network terminal power management system according to claim 1, characterized in that: When using finite-depth search as a fallback strategy, the search depth is limited by a predefined upper limit; candidate integer actions are scored based on a weighted cost function that includes switching penalties, indoor temperature deviation, and energy costs; the weights of the cost function are given by system configuration parameters; and the final action must meet the allowable range of indoor temperature and the upper and lower limits of the energy storage state of charge.

7. An embedded network terminal power management system according to claim 1, characterized in that: When a report is triggered, the reported data is differentially encoded and compressed. The differential encoding and compression includes three steps: transformation according to a preset sparse transformation, quantization using fixed-point quantization, and encoding in a differential manner. When privacy protection is required, differential privacy noise of a predetermined strength is injected. The encoded data stream is transmitted in an encrypted manner.

8. An embedded network terminal power management system according to claim 1, characterized in that: When communication is interrupted or distributed solution fails, the backoff control adopts simplified model predictive control or rule control: the simplified model predictive control uses linearized thermal balance constraints as the model, and transforms the optimization of continuous control variables into quadratic programming to generate feasible control sequences.

9. An embedded network terminal power management system according to claim 1, characterized in that: The processing unit also includes a data acquisition and preprocessing module, an online parameter identification module, a short-term prediction and fusion module, a rolling optimization and distributed solution module, a strategy mapping table query module, a rollback control module, and a communication and privacy module.

10. A control method for a power management system in an embedded network terminal, characterized in that: Receive the day-ahead operating trajectory from the microgrid central control unit and use the day-ahead operating trajectory as the initial solution for rolling optimization; Based on local observation data, online identification is performed to estimate building thermal model parameters, and a local optimization model for the current rolling period is constructed by linearization based on these parameters. Using any one of the following: long short-term memory neural network, autoregressive moving average model, or autoregressive model, and based on input features including historical indoor temperature, control actions, external disturbances, and time information, a local short-term forecast is generated. This local short-term forecast is then fused with correction information from the edge server to form a fused forecast for rolling optimization. Based on the initial solution, the local optimization model, and the fusion prediction, a local optimization subproblem is constructed. Compressed aggregated information is exchanged with the microgrid central control unit or a nearby terminal. The alternating direction multiplier method is used for distributed solution to obtain continuous variable candidate solutions within the limited computational quota. When the local optimization subproblem involves binary switch decisions, the policy mapping table is queried to obtain integer actions. If the distributed solution fails to reach the convergence criterion within the computational quota, the output of the policy mapping table or the result of the finite depth search is used as the backoff integer decision. In the event of communication interruption or local solution failure, rollback control is executed to ensure indoor comfort and energy storage safety; aggregated or compressed Lagrange information is reported in an event-triggered manner, and differential coding and compression coding are used to reduce communication volume; and event-triggered reporting is supported; the conditions for event-triggered reporting include any of the following: the absolute error between the local short-term forecast and the day-ahead operating trajectory exceeds a predetermined threshold; the equipment operating status is abnormal; the energy storage state of charge exceeds the preset upper and lower limits.

Images