Energy optimization model deployment framework based on center training-edge reasoning

Abstract

The invention provides a central training-edge reasoning energy optimization model deployment framework, relates to the technical field of artificial intelligence model deployment and edge computing, and adopts an event triggering, adaptive threshold and priority abstract mechanism to realize low-energy-consumption and efficient data management and model updating of edge equipment. The edge nodes continuously collect multi-source operation data, three types of core indexes including model performance degradation, energy consumption change and communication cost are generated through filtering and preprocessing, a comprehensive utility value is calculated and compared with a self-adaptive threshold value, and abstract uploading is decided by combining a probability triggering mechanism. The uploaded data are transmitted in a compact priority abstract mode, high-error samples, sparse gradient signatures and statistical characteristics are included, and low-bandwidth and high-representativeness data transmission is achieved. And the center side gathers multi-node abstracts, determines a model updating strategy through digital twinning and utility evaluation, generates a differential updating packet and safely issues the differential updating packet to edge nodes, thereby realizing multi-hardware adaptation, gray level updating and rollback guarantee.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence model deployment and edge computing technology, specifically to a center-training-edge-inference energy-optimized model deployment framework. Background Technology

[0002] With the rapid development of Industrial Internet, intelligent manufacturing, and Internet of Things (IoT) technologies, a large number of edge devices are performing real-time data acquisition and model inference on-site to achieve automated control, energy management, and intelligent decision-making. However, the existing edge computing and center training model has several technical problems that severely restrict the system's capabilities in terms of energy efficiency, stability, reliability, and real-time updates.

[0003] Problem 1: Edge devices are typically limited by computing power, storage, and energy supply. Deploying high-precision training or inference models directly at edge nodes can lead to excessive energy consumption, increased response latency, and elevated system thermal load. Especially in applications such as energy management and environmental control, frequent model updates and large-scale data uploads consume significant network bandwidth and increase the power consumption of edge nodes, thereby reducing overall operational efficiency. Existing methods largely rely on periodically uploading raw data to the center for model training or updates, lacking dynamic awareness and adaptive control of the edge node's operating status and network environment, which can easily cause network congestion, excessive communication energy consumption, and wasted computing resources.

[0004] Problem 2: Existing systems, when dealing with performance degradation of edge inference models, often rely on full data or coarse-grained metrics to determine whether an update is needed, making it difficult to accurately assess the model's actual performance on specific nodes or in specific scenarios. Because environmental variables, load conditions, and device status may change over time, the central system cannot promptly obtain local degradation information from edge nodes, leading to delayed or insufficient model updates, affecting control accuracy and energy efficiency optimization. Furthermore, the heterogeneity and distribution differences among edge nodes make it difficult for a single central system to adapt the model to the hardware and operational characteristics of all nodes, and traditional update mechanisms lack specificity and controllability.

[0005] Thirdly, existing model update schemes often ignore communication costs and energy consumption constraints. When edge nodes upload large-scale raw data or update model parameters, neglecting factors such as current bandwidth, network load, and electricity prices can easily lead to network congestion or energy waste during peak periods. Systems lacking opportunistic window scheduling mechanisms and adaptive upload strategies cannot effectively reduce transmission costs and struggle to achieve real-time model optimization under low-energy conditions. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a center-training-edge-inference energy optimization model deployment framework, which solves the following problems:

[0008] 1. Due to the limited computing power, storage, and energy of edge devices, directly deploying high-precision models can lead to high energy consumption, large response latency, and increased network load and system thermal load.

[0009] 2. To address the problem that existing systems struggle to promptly detect local model degradation and heterogeneous environments at edge nodes, leading to delayed model updates, poor adaptability, and impacting control accuracy and energy efficiency optimization;

[0010] 3. The model updates often ignore communication costs and energy consumption constraints, lack opportunistic window scheduling and adaptive upload strategies, which can easily lead to network congestion and energy waste.

[0011] Technical solution

[0012] To achieve the above objectives, the present invention provides the following technical solution: a center-training-edge-inference energy optimization model deployment framework, the framework comprising the following steps:

[0013] Sp1: Edge-side index calculation: The operational data collected by the edge nodes are used to calculate the first index representing the degree of model performance degradation, the second index representing the trend of energy consumption change, and the third index representing the communication cost. The three indexes are combined according to preset weights to obtain the comprehensive utility value.

[0014] Sp2: Adaptive threshold generation: Calculates the historical mean and fluctuation range based on the fluctuation of historical comprehensive utility values, and generates the adaptive threshold for the current node accordingly;

[0015] Sp3: Probability Trigger Judgment: When the comprehensive utility value is higher than the adaptive threshold, the synchronization probability is calculated based on the difference between the two, and the comparison between the synchronization probability and the random number determines whether to perform the summary information upload, so as to avoid concentrated uploads caused by short-term fluctuations.

[0016] Sp4: Priority Summary Generation: The collected data are prioritized based on factors such as anomaly degree, novelty degree, model output confidence and communication overhead. Several high-priority data are selected to form a summary package. The summary package includes sample feature statistics, model parameter change summary or sparse gradient change identifier.

[0017] Sp5: Opportunity Window Scheduling: When the emergency upload conditions are not met, multiple low-cost communication periods are identified based on network bandwidth prediction, electricity price prediction, and communication load prediction, and a summary packet is transmitted during the communication periods.

[0018] Sp6: Central side scheduling decision: The central side receives summary packets from multiple edge nodes, calculates the node utility value based on the model degradation information, energy consumption change information and communication consumption information contained therein, and determines the set of edge nodes that need to be updated under the central resource budget constraints.

[0019] SP7: Differential Model Distribution: The central side sends the differential update content of the training model to the corresponding edge nodes according to the scheduling results, and generates multiple versions of quantized models in combination with the hardware type of the edge nodes to improve the energy consumption efficiency and execution efficiency of the inference stage.

[0020] Preferably, the edge-side index calculation is based on the output error change of the continuous time window data calculation model, the sampling change of device energy consumption, and the estimated value of communication overhead to generate the first index, the second index, and the third index.

[0021] Preferably, the adaptive threshold generation adopts an exponential moving average or an adaptive update rule based on historical utility value fluctuations to generate the adaptive threshold, so that the triggering conditions automatically converge or relax as the node's operating status changes.

[0022] Preferably, the probability triggering determination is based on the magnitude of the comprehensive utility value exceeding the adaptive threshold to construct a monotonically increasing probability, so that the higher the comprehensive utility value, the greater the corresponding synchronization probability.

[0023] Preferably, the priority summary generation performs scoring based on error value, feature change degree, model output confidence and upload bandwidth cost, and selects data to form a summary package within the target size limit.

[0024] Preferably, the summary package includes at least one of the following: statistical features of high-loss samples; sparse differencing of model weight changes; a compact statistical structure of feature anomaly degree; and an event index structure for abnormal periods.

[0025] Preferably, the opportunity window scheduling determines multiple potential low-cost transmission periods based on future bandwidth prediction, electricity price prediction, or communication energy consumption prediction, and performs non-urgent digest packet transmission within the specified periods.

[0026] Preferably, the central-side scheduling decision selects the update node that meets the energy consumption and bandwidth limitations by sorting the node utility values ​​or using integer programming under resource budget constraints.

[0027] Preferably, the differential model distributes model update content in a differential manner and generates multiple versions of quantized models based on the hardware type of the edge nodes to reduce energy consumption during the model update and inference processes.

[0028] Beneficial effects

[0029] This invention provides a center-training-edge-inference energy-optimization model deployment framework. It offers the following advantages:

[0030] 1. This invention employs lightweight data preprocessing and local statistical analysis at edge nodes, avoiding direct high-precision training at the edge, thereby significantly reducing computational and energy consumption pressure. Edge nodes utilize low-latency algorithms such as average filtering and median filtering to denoise and smooth multi-source data, generating statistics for index calculation. Simultaneously, a high-information-density priority summary is constructed locally, uploading only key samples and sparse parameter variations. After compression and quantization, this is transmitted through a secure channel. This reduces large-scale data backhaul while ensuring data parsability and integrity, avoiding response latency and system heat load caused by frequent uploads and large-volume communication, ensuring stable and efficient operation of the edge inference model even under energy-constrained conditions.

[0031] 2. This invention employs local calculation of fine-grained model performance degradation indicators, energy consumption changes, and communication costs at edge nodes, and generates a comprehensive utility value that is dynamically compared with an adaptive threshold to determine the necessity of uploading. Combined with a probability triggering mechanism and an emergency triggering channel, it achieves precise control over uploading behavior, avoiding the lag and insufficiency of relying solely on coarse-grained indicators or full data. Simultaneously, the central side generates multiple adapted model versions based on the node hardware configuration, including quantization, pruning, distillation, and accelerator-optimized versions. Then, through differential updates and canary release mechanisms, it achieves targeted, safe, and controllable model updates, enabling edge nodes to maintain high-precision control and energy efficiency optimization under heterogeneous hardware and diverse operating conditions.

[0032] 3. This invention employs an opportunistic window scheduling mechanism and bandwidth, electricity price, and communication load prediction to optimize the timing of summary packet uploads, enabling the upload of low-cost, high-priority samples first. Simultaneously, differential compression and block transmission are used to reduce the amount of data transmitted, and only lost segments are retransmitted in case of transmission failure, reducing energy consumption and the risk of duplicate transmissions. On the central side, combining node utility values ​​and resource constraints, integer programming or approximate algorithms are used to select the optimal node combination for model distribution, achieving maximum benefit updates under limited bandwidth, energy budget, and computing resources. This effectively avoids network congestion and energy waste during peak periods, realizing real-time, low-energy model optimization through edge-center collaboration. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the deployment framework of the energy optimization model of the present invention.

[0034] Figure 2 A framework diagram of the energy optimization model deployment framework of the present invention;

[0035] Figure 3 This is an architecture diagram of the energy optimization model deployment framework of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1:

[0038] like Figures 1 to 3 As shown, a center-training-edge-inference energy optimization model deployment framework is presented, which includes the following steps:

[0039] Sp1: Throughout the system operation, edge nodes continuously collect operational data through local sensor interfaces, device communication bus interfaces, network interfaces, and system monitoring modules. This includes physical quantities provided by temperature and humidity sensors, fan tachometers, compressor current and voltage probes, smart meters, and ambient light sensors; device information such as the on / off status, gear parameters, and error codes of device controllers; external environmental variables provided by outdoor meteorological APIs or local meteorological modules; throughput, latency, and packet loss rate of Ethernet or wireless link interfaces; real-time electricity prices provided by electricity price service interfaces or local energy storage systems; and edge-side computing power and energy consumption status reported by internal system monitoring modules, such as CPU utilization, operating temperature, and battery level. All data is collected at specified intervals, such as equipment operation data using intervals of one to ten seconds, cumulative energy consumption and electricity price information using intervals of more than one minute, and environmental variables and link quality using self-adjusting sampling intervals based on fluctuation amplitude.

[0040] The acquired data is input into the local preprocessing module of the edge node. This module eliminates abnormal spikes and sensor noise through built-in moving average filtering, median filtering, or other low-latency filtering algorithms, and performs unified timestamp correction on multi-source data to ensure comparability of data from different interfaces at the same time. When using averaging filtering, each time the edge node receives new data from a sensor, it places that data along with historical data from the same sensor within the most recent sampling periods into a fixed-length sliding window. The system sequentially reads all sampled values ​​from this window, sums these values, and then divides the sum by the number of data points in the window to obtain a new smoothed value. This smoothed value serves as the "denoised data" for the current moment, replacing the original single-sample value. In this way, if a sudden spike occurs in a single sampling period, the window averaging will smooth out the impact of this abnormal value, making the obtained smoothed value closer to the sensor's true trend. The window continuously slides forward as new data arrives, adding the latest data and removing the oldest data, ensuring that the smoothed value is always based on the actual operating conditions over the most recent period. The output of the average filter is written to the time-series database on the edge side, allowing subsequent calculation modules such as performance degradation indicators and energy consumption change indicators to read it directly, thus ensuring the stability and reliability of the data used by subsequent algorithms. When median filtering is used, after the edge node receives new sensor data, it combines this data with the most recent sampled values ​​within the window to form a numerical set. The system rearranges this set in ascending order of numerical value, and then selects the value in the middle position of the rearranged set as the filter output for the current moment. If the window size is odd, the middle value is taken; if the window size is even, the average of the two middle values ​​is taken as the output. This processing method can effectively suppress single or small noise spikes, such as isolated outliers caused by instantaneous sensor jitter or communication errors, because outliers are placed at the beginning or end of the set after sorting, without affecting the true data in the middle position of the set. Therefore, the value output by the median filter usually represents the "typical level" within the current window, and has a stronger resistance to outliers. The smoothed data generated by the median filter is also written to the edge preprocessing database for subsequent indicator calculations, and is particularly suitable for sensor data with high noise or occasional distortion. When data is missing or arrives late, the preprocessing module uses interpolation mechanisms or a previous value preservation strategy to fill in the missing data, ensuring the continuity and stability of subsequent calculations. Simultaneously, the preprocessing module calculates statistics such as average, variance, extreme values, and growth rate according to a preset time window, and stores the raw data and statistics in a local lightweight time-series database for access by the real-time indicator calculation and summary construction modules.

[0041] After obtaining the preprocessed data, the edge nodes proceed to calculate three core metrics. First, a primary metric representing the degree of model performance degradation is generated. This metric originates from the deviation between the control commands or predicted values ​​obtained by the model after inference at the edge and the actual feedback. For example, in a temperature control scenario, the average value, rate of change, and long-term offset of the error between the model's predicted temperature and the sensor's measured temperature over the most recent few minutes are all included in the calculation. If a reinforcement learning model is used, the performance degradation trend can also be measured by the recent decrease in reward values. This metric expresses the degree of decline in the model's adaptability to the current scenario. All error information is directly generated from the model's inference output and real-time sensor readings and written to a local cache for subsequent use by the comprehensive utility value generation module. The second metric characterizes the energy consumption change trend. The edge nodes read real-time power and cumulative energy consumption data from the smart meter interface, compare the power sequence generated by the preprocessing module with the historical baseline power sequence, and estimate the magnitude of energy consumption change by combining the current operating status of the equipment (such as the number of compressor starts and fan speed changes). To improve predictive capabilities, the system maintains a lightweight energy consumption baseline model at the edge. This model compares the average energy consumption calculated from historical data for similar time periods (e.g., the same moment, the same load level) with the current energy consumption to determine if the current energy consumption exceeds expectations. This metric is used to evaluate whether the current model control strategy may lead to increased energy consumption, thus providing a quantitative basis for whether to trigger model synchronization at the energy consumption level. A third metric is used to estimate communication costs. Edge nodes read actual bandwidth, real-time latency, and packet loss rate information from the network interface and estimate the communication time required for one upload operation and the possible number of failed retransmissions based on the current network state and digest packet size. In wireless scenarios, transmission energy consumption is also estimated based on transmit power and transmission duration. Communication costs can be expressed as the expected amount of data transmitted, expected energy consumption, or converted monetary costs, specifically determined by the edge node's configuration based on the scenario strategy. A higher value for this metric indicates a greater cost for model synchronization under the current network conditions.

[0042] Once all three types of indicators have been calculated, the system combines them linearly or non-linearly according to preset weights to form a comprehensive utility value. The weights can be adjusted periodically by the central side through the strategy distribution module, or adaptively updated by the edge side based on long-term trends. The comprehensive utility value expresses the relative relationship between the benefits and costs of performing a single model synchronization under the current operating state. The higher the comprehensive utility value, the greater the model degradation, the greater the energy-saving potential, and the lower the communication cost. The comprehensive utility value is not only used for synchronization trigger judgment but is also saved to the historical sequence cache of the edge nodes for the adaptive threshold generation module to analyze long-term trends.

[0043] Sp2: In edge nodes, the adaptive threshold is generated based on historical aggregate utility value sequences. Edge nodes first locally store historical aggregate utility values ​​for a certain time period, such as continuous data from the most recent 24 hours or 7-day multi-24-hour data recorded weekly. This historical data is obtained from each aggregate utility value calculation and stored in the edge database with timestamps. Through statistical analysis of the historical sequence, the node first calculates the historical mean, which is obtained by summing historical aggregate utility values ​​over a specified time window and dividing by the sample size. Then, the volatility is calculated by taking the square root of the average of the squared differences between each historical value and the historical mean to obtain the standard deviation. The standard deviation is used to quantify the magnitude or instability of aggregate utility value changes. The node combines the historical mean and volatility to form the adaptive threshold. The basic formula is: Adaptive threshold equals historical mean plus a conservative factor multiplied by volatility, where the conservative factor is a coefficient greater than zero used to control the threshold's sensitivity to volatility. A conservative factor can be set to a larger value during the initial deployment phase to reduce false triggers caused by short-term fluctuations. As the system runs and historical data accumulates, it can be gradually lowered based on actual performance to ensure the threshold more accurately matches the node's current operating state. To enable the threshold to dynamically follow the trend of the comprehensive utility value, nodes can use an exponential moving average method, assigning higher weights to recent comprehensive utility values ​​and lower weights to earlier data. Specifically, the current comprehensive utility value is multiplied by a specified smoothing coefficient, and the previous threshold is multiplied by one minus the smoothing coefficient, thus generating a new adaptive threshold. This method can smooth out short-term fluctuations caused by sudden anomalies, allowing the threshold to adjust gradually with the operating state without being excessively raised or lowered by a single anomaly. Nodes periodically calculate and update the adaptive threshold, for example, after each comprehensive utility value calculation, to ensure the threshold remains synchronized with the current system state. The generated adaptive threshold and its version information are stored in an edge database for direct retrieval during subsequent upload decisions or scheduling processes. The update timestamp and parameter information are also recorded for auditing and backtracking. In this way, the threshold can be automatically tightened or loosened as the node's operating environment changes over a long period of time, maintaining a reasonable triggering level in different seasons, load cycles, or network conditions. This avoids concentrated uploads or resource waste caused by periodic fluctuations or short-term anomalies, thus effectively supporting subsequent probability triggering and opportunity window scheduling strategies.

[0044] Sp3: After the edge node completes the calculation of the comprehensive utility value and obtains the adaptive threshold, the system first compares the magnitudes of the two to determine whether the current operating state deviates from the long-term stable range. The comprehensive utility value comes from the edge-side indicator calculation module and is composed of three indicators: the degree of model performance degradation, the trend of equipment energy consumption change, and communication cost, combined according to preset weights. It is used to reflect the current comprehensive operating state of the node. The adaptive threshold comes from the mean and fluctuation range of the comprehensive utility value in the historical time window, which can describe the typical fluctuation range of the node under normal operating conditions. When the comprehensive utility value exceeds the adaptive threshold, it indicates that the current model prediction error, energy consumption level, or communication load has deviated abnormally, and the system needs to further determine whether summary data should be uploaded.

[0045] To quantify the degree of deviation, edge nodes first calculate the difference between their comprehensive utility value and an adaptive threshold. This difference, obtained by direct subtraction, measures the deviation of the current state from the normal fluctuation range. A larger difference indicates a more significant decline in model performance, a greater increase in device energy consumption, or more strained communication resources, suggesting a more severe abnormal state for the node. A small difference represents only a minor disturbance; however, a significantly increased difference indicates that the system state has clearly deviated from the historical normal range. The system does not immediately initiate an upload simply because the difference is greater than zero. Instead, it inputs the difference into the probability mapping module. This module uses a continuous, monotonically increasing smooth mapping rule to convert the linear quantity of the difference into a synchronization probability between zero and one. When the difference is small, the mapped synchronization probability is close to zero, indicating that the node state has only undergone a minor change, and the necessity of uploading is low. As the difference gradually increases, the mapped synchronization probability increases accordingly. When the difference reaches a high level, the mapping result approaches the maximum value, and the node is considered to be in a strongly abnormal state, requiring the upload of summary information as soon as possible so that the central side can obtain crucial information. After obtaining the synchronization probability, the edge node generates a pseudo-random number between zero and one and compares this random number with the synchronization probability. If the random number is less than the synchronization probability, the system determines that the current anomaly level has reached the upload condition, and the node immediately enters the summary generation process and prepares for data transmission. If the random number is greater than the synchronization probability, this anomaly event is considered a minor fluctuation, and the node will write the relevant information of the event into the evaluation list and continue to observe the change in the overall utility value in the next time window. This avoids multiple nodes uploading simultaneously due to short-term disturbances, which could cause network congestion and energy consumption peaks.

[0046] To handle critical situations related to model stability or production safety, the system also sets up an independent emergency trigger table. This trigger table relies on abnormal signals provided by the device's monitoring module, such as hardware temperature exceeding limits, self-test failure, distorted readings of key sensors, and alarm codes from production equipment. Once these situations are detected, the node bypasses the entire probability determination process and directly enters the forced upload channel, ensuring that the central side can obtain the critical status immediately and execute scheduling or model repair in a timely manner. Through this overall mechanism, the upload behavior of edge nodes is no longer fixed and periodic, but simultaneously possesses three capabilities: suppressing unnecessary uploads caused by short-term fluctuations through probability mapping; increasing upload enthusiasm when the anomaly level increases through the correspondence between difference and probability; and immediately uploading in serious safety-related situations through the emergency channel. In addition, the addition of pseudo-random numbers makes the upload times of different nodes under similar anomaly levels more decentralized, effectively reducing communication congestion caused by concurrent uploads and improving the stability and energy efficiency of the entire edge cluster.

[0047] Sp4: After the edge node completes the aforementioned comprehensive utility value determination and triggers the upload condition, it does not perform a large-scale backhaul of the original data. Instead, it constructs a priority summary with high information concentration and extremely low communication load locally. This process relies on the real-time data buffer, feature extraction module, model forward inference module, and error recording module within the edge node.

[0048] During continuous operation, edge nodes store sensor-collected data, device operating status data, task input features, and model prediction outputs in a local short-term buffer in chronological order. This buffer typically uses a circular structure, retaining only data from the most recent period, such as the last few minutes or hundreds of records. Each record includes the original input features, the model's inference result, the actual observations from the device, inference latency, energy consumption sampling during inference, and the system's confidence output. All of this data is used as candidate samples for priority evaluation during summary generation. Before generating the summary, a priority score is calculated for each candidate sample in the buffer. The priority score is composed of multiple factors, including the instantaneous error generated by the sample, calculated as the difference between the model's inference output and the actual observations, which characterizes the degree of local degradation of the model. Feature novelty is measured by the distance between the current sample's features and the existing feature distributions in the previous training set. This distance can be obtained directly from the local feature cache by the directional difference of feature vectors, the distance of cluster centers, or other reproducible similarity calculation methods. Samples with higher novelty are more representative. The model's output confidence score for this sample is directly provided by the model inference module. The lower the confidence score, the less certain the model is about the input, and such samples need to be uploaded so that the central node can correct the model's performance in that input region. Furthermore, the edge node locally calculates the communication load required to upload the sample. This communication load, combined with the current bandwidth status, packet size estimate, and expected transmission energy consumption, allows nodes to suppress uploads when bandwidth is limited or energy consumption is high, thus achieving a balance between uploaded data and resource constraints.

[0049] The priority scoring module normalizes error, novelty, confidence, and communication cost separately, and then combines them into a final score according to preset weights. After all candidate samples are sorted from high to low according to this score, a small subset of samples with higher scores are selected as the objects for summary construction. The number of samples selected is usually several to dozens, and the choice of number depends on the current bandwidth and the target maximum summary size limit, so that the final summary is both representative and can be stably transmitted in low-bandwidth channels.

[0050] To reduce communication overhead, edge nodes do not upload the complete raw data packets. Instead, they construct a compact description for each selected sample. This compact description includes a feature summary, quantized feature values, a timestamp, model inference results, true observations, error values, and a sparse signature structure to characterize the local parameter variation trends. The feature summary is a set of features selected from key sample features, such as those with the most significant gradient changes or those highly correlated with error. These features are further quantized, compressing floating-point numbers into a smaller data representation, thus reducing space usage. The sparse signature structure represents the parameter dimensions that change most significantly when the model processes these samples. These dimensions are filtered by a threshold, recording only the indexes of parameters exceeding a specified amount of change and their corresponding parameter differences. This allows the central node to determine from this sparse information which local regions the model has shifted, without transmitting the entire gradient or weight parameters.

[0051] In addition to sample-level summaries, nodes also generate statistical summaries to describe the overall data structure within the buffer. For example, they calculate the histogram distribution of model inference confidence, the latency distribution of inference execution, and the distribution of sampling energy consumption. These statistics reflect whether there are systematic deviations in the node's current operating state, such as an overall decrease in confidence or an overall increase in energy consumption. If a node detects an abnormal time period, such as a period of continuous error increase, it records the event index for that time period. The location of these abnormal time periods is recorded using a compact tagging structure, such as a counting-type data compression structure, allowing the central side to determine the distribution of problems in the model over time without needing to obtain the specific raw data.

[0052] The completed digest package includes node number, digest generation time, number of samples in the digest, key features and quantization values ​​of each sample, sample error, sample timestamp, sparse gradient index and difference values, statistical histogram data, energy consumption statistics, and a digest hash fingerprint used to verify the integrity of the digest package. All information is compressed and encoded into a small binary serialization format by the serialization module, for example, using a lightweight binary encoding and decoding format, enabling stable transmission of the digest package in typical low-bandwidth industrial edge computing networks. To ensure security and integrity, the digest package undergoes integrity signing and encryption processing by a local signing and encryption module before final transmission. It is then sent to the central server via the edge node's communication module using message queue transmission or remote call transmission, resulting in a digest with extremely high compression ratio, parsability, and security.

[0053] Sp5: When a digest packet is generated at the edge node but no emergency upload is triggered, the digest packet is stored in a local opportunity window queue, which is maintained by the scheduling module within the edge node. The scheduling module continuously receives input from multiple data sources, including current uplink bandwidth samples collected in real time by the bandwidth monitoring module, historical bandwidth sequences recorded by the network interface driver layer, electricity price data from local energy metering devices or the upper-level energy management system (if electricity prices are available in real time), and recent communication load changes statistically analyzed by the network load monitor. Based on this data, the scheduling module constructs a prediction of communication costs in the near future to select a low-cost upload opportunity.

[0054] Bandwidth prediction is handled by a lightweight bandwidth prediction submodule. This submodule periodically retrieves bandwidth sample values ​​from the network interface for several past time slices, performs an autoregressive calculation on these samples, and generates a prediction of the potential bandwidth size for the next time slice by calculating the weighted relationship between the sample values ​​from the previous time slice and several past samples. This prediction describes the available bandwidth that may be achieved during future transmissions; higher bandwidth indicates lower unit data cost, making it more suitable for upload periods. Simultaneously, if electricity price data is available in the system, it is obtained by reading the real-time electricity price table provided by the power system or a locally set electricity price curve to determine the electricity price level for several future time slices. Time slices with lower electricity prices are marked as "energy cost reduction periods." The scheduling module also integrates communication load data, which is statistically analyzed by the network stack, such as average uplink traffic, number of connections, and congestion indicators for each time slice. This data allows for the estimation of peak and off-peak periods; lower load corresponds to a lower probability of congestion, thus reducing energy consumption and the probability of duplicate transmissions.

[0055] After obtaining bandwidth, electricity price, and communication load forecasts, the scheduling module integrates all forecast results using a unified cost model. The cost model maintains a cost calculation rule that combines the bandwidth, electricity price, and communication load forecasts with preset weights to derive the total communication cost score for each future time slice. Higher bandwidth results in a lower cost score, higher electricity prices result in a higher cost score, and higher communication loads also increase the cost score. The cost calculation results form a time slice sequence, which is sorted from lowest to highest cost. Several low-cost time slices are selected as potential opportunity windows.

[0056] When a digest packet is in a waiting state, the scheduling module checks whether it has entered one of these windows within the calculated low-cost time slice. When the current time slice falls into a selected window, the scheduling module retrieves digest packets from the opportunity window queue in priority order for transmission. Digest packets already have a priority score when entering the opportunity window queue, calculated by the digest generation module based on factors such as data anomaly severity, model output uncertainty, and communication costs. Regarding the transmission order, the scheduling module starts by selecting the highest priority digest packet. If the remaining bandwidth is sufficient to support the upload of multiple digest packets, they are uploaded in batches from highest to lowest priority. If the bandwidth prediction module indicates limited available bandwidth, the scheduling module automatically reduces the upload batch size, prioritizing high-impact digests to prevent model degradation due to delays. To further reduce transmission energy consumption, digest packets are handed over to the edge-side data compression module for differential compression before uploading. The differential compression module compares the digest packet with a snapshot of the digest structure or model parameters from the previous upload, retaining only the feature statistics, weight difference information, or anomaly event indexes of the changed parts, thereby reducing the final uploaded data volume. If the digest packet size is still large, the scheduling module divides the digest packet into multiple data blocks based on the bandwidth prediction of the current time slice. Each block is accompanied by a sequence number for reassembly at the central side. Edge nodes upload these data blocks in sequence via network drive. In case of transmission failure, the corresponding segment can be resent according to the block sequence number, reducing the energy consumption of repeated transmissions. When a sudden improvement in network conditions is detected, such as when the bandwidth prediction module identifies a sudden increase in current bandwidth, the scheduling module allows more blocks in the queue to be sent immediately in advance, thereby completing more digest transmissions under high bandwidth conditions and further reducing overall communication energy consumption.

[0057] After uploading, the summary packet is removed from the opportunity window queue and sent to the central server's receive buffer via the transmission interface, serving as input for subsequent central model scheduling decisions. Ultimately, the center uses this summary data to determine whether model updates are needed for that node or multiple nodes, thus forming a closed-loop optimization process.

[0058] SP6: After receiving summary data from multiple edge nodes, the central platform immediately writes each summary packet into the central data lake storage structure. The data lake records the source node, collection time, data type, and statistical features or sparse differential content contained in the summary through a metadata directory, ensuring accurate traceability for subsequent analysis. After the summary enters the data lake, the central scheduling module first uses the model degradation information, energy consumption change information, and communication overhead estimates contained in the summary to perform a lightweight analysis of the current node's status. This analysis uses the central local monitoring database, model version records, and node historical performance as auxiliary information, enabling the center to quickly identify whether a node is at high risk of degradation or has excessive energy consumption without needing complete data. Based on this, the central evaluation module calls the local digital twin inference system or the built-in simulation module to generate potential improvements after model updates by simulating the node's operating status, inference pressure, data distribution drift probability, and energy consumption curve. The inputs to a digital twin system include feature statistics from the current summary, node hardware specifications, the last model update time, running power samples, and communication records. The outputs are estimates of potential reductions in inference energy consumption, inference errors, and improved response latency after the model update.

[0059] The center then calculates the node utility value for each edge node, which consists of three parts: the first part is the model degradation magnitude, calculated based on the error growth, confidence decrease, or weight drift recorded in the summary; the second part is the potential energy consumption benefit, composed of the expected energy savings, expected improvement in inference stability, and reduction in model inference time derived from the digital twin inference results; and the third part is the synchronization cost, generated based on the node's reported communication bandwidth usage, the cost of uploading the summary, and historical communication cost estimates. These three parts are combined according to the strategy weights set by the center, enabling the center to measure whether each node is worth updating its model in this round within a unified evaluation system. A higher node utility value indicates a greater benefit from updating the node when synchronization resources are limited.

[0060] After calculating the node utility values, the central scheduling module checks the current central resource budget, including the bandwidth limit available for model distribution, the energy consumption budget for push from the center to the edge, and the number of model compilation resources available on the central side. Based on this real-time resource information, the center constructs an optimization selection process, expressed using an integer programming model. The decision variables in the integer programming represent whether each node is in the current update list. The objective is to maximize the sum of the utility values ​​of all selected nodes while satisfying the bandwidth budget, energy consumption budget, and central computing capacity constraints. The constraints of the integer programming include the bandwidth cost required for each node to distribute the model, the central computing resource limit when multiple nodes update in parallel, and the peak load limit of the central network. The center's built-in solver module (such as OR-Tools, GLPK, or CBC) can solve for the optimal node combination under these constraints.

[0061] When the number of edge nodes is large, the central scheduling module automatically switches to an approximate solution to avoid excessively long solution times. This approximation method sorts nodes by their utility-to-cost ratio and selects them one by one from high to low until any resource limit is reached, forming a near-optimal set of nodes. The final set of nodes is recorded in the central model version control system, which generates update records, including the basis for selecting a node, its current utility value, resource budget status, and summary content. It also generates corresponding audit trails for subsequent tracking and verification. After this process, the central scheduling module delivers the results to the backend model distribution module, enabling the center to update the most valuable edge nodes within a controllable cost. It ensures that every update decision is interpretable, traceable, and auditable, thus maintaining a stable, efficient, and intelligent update rhythm for the entire central training-edge inference system under energy-constrained conditions.

[0062] Sp7: After completing a new round of model training or receiving summary packets uploaded by several edge nodes, the central system aggregates these packets to form a new model version. To ensure the model can adapt to different types of edge nodes, the central system first constructs multiple model variants based on the hardware configuration files of each edge node. The hardware configuration files are reported by the edge nodes upon initial system access and include information such as processor architecture, whether a dedicated acceleration unit is available, available memory size, computing power level, and energy consumption limitations. Based on these hardware characteristics, the central system generates different versions of the trained model, including but not limited to a full-precision version for high-performance devices, a structured pruning version for general embedded devices, an eight-bit quantized version for low-power devices, an optimized version for TensorRT or EdgeTPU for dedicated accelerators, and a distilled lightweight version for extremely constrained hardware. Each model version undergoes verification and inference after generation to ensure that model structure transformation and quantization do not affect basic inference correctness. To reduce communication overhead during model updates, the central system compares the new model with the old models currently stored on each edge node and calculates the weight differences between them. These differences stem from the binary content of the model files. Therefore, the center employs a method similar to incremental file synchronization algorithms, generating differential update packets by analyzing the different portions of the binary blocks in the two model files. This differential packet only contains changed weight segments, pruning information, or changes in quantization parameters; unchanged portions are not retransmitted. After generation, the differential packet is compressed to reduce file size and then digitally signed by the center's security module to ensure that edge nodes can verify the packet's origin and integrity upon receipt.

[0063] Once the differential model and quantized version are ready, the center determines the distribution order to each edge node based on the scheduling decisions made at the center. The scheduling process comprehensively considers the center's current bandwidth budget, overall energy consumption constraints, and predicted network status for each node. For example, updates are performed on more nodes during periods of lower communication costs, while only necessary nodes are updated during periods of high load. The center sends the generated differential update packets to the corresponding edge nodes through the system's message transmission module. The message transmission module sends data through a secure channel and monitors the transmission status.

[0064] Upon receiving the differential update packet, the edge node first verifies the digital signature and checksum through the verification module to confirm the integrity and authenticity of the data source. After successful verification, the edge node prepares two model storage locations in its locally maintained model management module: the currently stable model version and a temporary storage area for the model to be updated. The differential packet is applied to the temporary model area, and the received incremental content is combined with the locally saved old model file using a differential synthesis algorithm to generate a complete new model version. If data inconsistency, missing fragments, or verification failures occur during the synthesis process, the node will immediately stop updating, retain the old model, and send an error report to the center. If the synthesis is successful, the node will perform a self-checking inference process in the isolated environment. The self-checking samples are pre-embedded in the model structure by the center or distributed with the version update to check for issues such as unresolved operators, dimension mismatches, or numerical anomalies.

[0065] Only when the self-check inference result is normal will the new model replace the old model and be added to the node's actual inference process; otherwise, the node will automatically roll back to the previous stable version and report the anomaly to the center. The center will decide whether to reissue the fixed differential package or postpone the node's update based on the error information. To further ensure the stability of large-scale systems, the center also has a canary update function, that is, to prioritize the trial of the new model version on a small number of nodes, and gradually expand the update scope after confirming that there are no large-scale compatibility issues, so that the overall system remains stable during model iteration. Through the above methods, the center can automatically generate adapted model versions for different hardware nodes, and significantly reduce the communication volume required for model updates through differential methods, while ensuring that the edge side can safely, reliably, and rollbackably complete the model replacement during the update process, thereby achieving low-energy continuous model optimization capabilities in the context of collaborative training at the center and inference at the edge. Specific Implementation Example 2:

[0067] like Figures 1 to 3 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0068] Within the energy optimization system that integrates central training and edge inference, the central side is responsible for model building, training optimization, compression, and version management, while the edge side is responsible for real-time inference, device control, data summary generation, and secure rollback. To adapt to different task requirements in the energy system, including load forecasting, local energy consumption regulation, device control strategy selection, and multi-device collaborative optimization, the central side builds various types of models and achieves accurate training through multi-source data. The types of models built by the central side mainly include: supervised regression models for numerical prediction, sequence prediction models for time series modeling, reinforcement learning policy networks for action decision-making, and ensemble models for robust fusion. To enhance the reproducibility and applicability of the model system, all types of models adopt structured methods that can be implemented under existing technological conditions and have clear deployability.

[0069] The supervised regression model employs shallow or medium-sized multilayer fully connected networks, random forests, or gradient boosting trees as its main structure, combined with an embedded feature encoder to handle discrete features. Inputs include current real-time sensor readings, equipment operating range, environmental variables, periodic time features, sliding statistical features, average load over the past few minutes, and equipment operating modes. To enhance model robustness, a feature attention weighting module is added within the regression model, dynamically assigning weights to input features based on different operating conditions. Model outputs include power prediction, temperature deviation prediction, and short-term energy consumption prediction, with confidence scores calculated heuristically or Bayesianly to guide local energy consumption estimation and policy screening for edge devices.

[0070] The sequence prediction model employs gated recurrent units, long short-term memory networks, or lightweight Transformer structures as its foundation. Internally, it incorporates a positional encoding compression module and a seasonal trend decomposition unit, enabling it to identify the cyclical, seasonal, diurnal, and rapidly changing burst load trends of equipment operation. Furthermore, to enhance its adaptability to complex sequences, the model features a multi-step prediction head structure: one responsible for short-term, strongly constrained predictions, and another for long-term trend predictions. Together, they form the prediction interval, providing multi-scale forward-looking input for subsequent strategy models.

[0071] The reinforcement learning policy network employs multi-head perceptrons, shallow convolutional networks, or lightweight Actor-Critic architectures as policy and value function evaluators. To enhance reproducibility, policy training utilizes a large number of simulated trajectories generated in a digital twin environment, based on temporal difference learning, proximal policy optimization, dual-delay policy gradient, or soft Actor-Critic (SAC) algorithms. The policy network's input includes future load predictions from the sequence model, current device state, action constraint boundaries, and locally adjustable parameters. The output includes control action suggestions (such as valve opening adjustment, power allocation, and task scheduling mode selection) and policy confidence, used by edge devices to execute under safety constraints. To avoid policy overfitting, a policy smoothing mechanism is also incorporated to ensure policy stability even with noisy inputs.

[0072] Ensemble models are used to improve overall stability and generalization under complex conditions. The ensemble system includes three types of base models: master regression models, master sequence prediction models, and lightweight calibration models. Through a weighted fusion layer, the outputs of each model are weighted with uncertainty awareness. The weights are determined by the following factors: the model's historical error statistics, the current input domain offset, the output variance, and the stability level of the model version. Furthermore, to avoid information redundancy between models, the ensemble model includes a decorrelation module. Through random sub-features, Dropout Bagging, or few-sample perturbation training methods, the base models are differentiated, thus maintaining the robustness of the overall output even when one model deviates.

[0073] The data sources used to train the above models include three categories: First, summary packages uploaded by edge nodes on an event or periodic basis, including high-error sample statistics, sequence feature summaries, model confidence histograms, power peak records, small window fluctuation statistics, and abnormal scene markers; second, real samples uploaded in batches after on-site anonymization, with anonymization steps including feature scaling, category remapping, sensitive field removal, noise injection, and differential privacy perturbation, ensuring that the data retains feature relationships without disclosing sensitive information; and third, synthetic trajectories generated by the digital twin system, which are generated by a hybrid of equipment physical models (such as thermal balance equations, power curve models, and start-up / shutdown inertia models) and data-driven models (such as autoregressive models and deep sequence models), ensuring that the training data covers long-tail scenarios, extreme operating conditions, and abnormal control behaviors. The central training process begins with data cleaning, including outlier removal, missing value imputation, time series resampling, synchronous calibration, feature normalization, and robust scaling. Subsequently, feature engineering is performed, using sliding window construction, Fourier time features, seasonal offset encoding, and physical derivation variables of electrical equipment (such as current square integrals and threshold event counts) to improve the quality of the input space. During training, hyperparameter search, cross-validation, and time-slice validation are performed on each model to ensure model stability across different operating cycles (peak hours, low-load nighttime periods, and seasonal transitions). After model training, three types of compression are performed: structured pruning, precision-aware quantization training, and knowledge distillation, enabling the model to maintain high accuracy even with limited computing power on edge devices. After compression, the model is stored in a model repository along with complete metadata, including the adapted hardware platform (CPU / GPU / NPU), inference latency estimate, maximum memory usage, model version number, quantization bit width, model validation fingerprint, recommended threshold range, and historical stability level. Edge nodes automatically select the most suitable model version based on their own hardware description.

[0074] When running at the edge, all models can output their own uncertainty metrics. For regression and sequence models, deep stochastic forward propagation, multi-head output variance, or Monte Carlo Dropout are used; for ensemble models, uncertainty is quantified by differences in output between models; for reinforcement learning strategies, confidence is assessed through value bias and policy entropy. When uncertainty exceeds a threshold, indicating a lack of confidence in the current situation, the edge node immediately switches to a safe fallback strategy. This fallback strategy consists of a rule engine, an edge MPC controller, or a lightweight linear model, ensuring the device operates within a safe range. The edge node simultaneously constructs a high-granularity summary and uploads it via an emergency channel, enabling the center to identify potential abnormal input domain offsets the model may face. If the edge model signature verification fails or internal checkpoints are corrupted, it automatically rolls back to a stable version and logs the audit. The edge node caches relevant data in a local table structure, including the original sampling table, summary queue list, and model version table. The summary package is serialized using Protobuf and uploaded to the central API gateway via an MQTT encrypted channel or gRPC. The center's training scheduling service writes the summary content to object storage and a database for use by the training module. The model repository supports differential generation, allowing edge users to download only the differences from the previous version, thus saving bandwidth resources. The system's data flow path includes: data acquisition and preprocessing, local inference and control decision-making, summary construction and uploading, central training and evaluation, model compression and version release, and edge validation and deployment. The summary package structure includes node identifiers, hardware descriptions, current model version, time window range, feature statistics, high-error sample summaries, sparse gradient indexes, confidence histograms, average and peak power, and integrity signatures. The differential package contains the base version number, target version number, differential content, compressed size, and signature, used for complete model recovery. Specific Implementation Example 3:

[0076] like Figures 1 to 3 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0077] The center-training-edge-inference energy optimization model deployment framework forms a continuous, adaptive, and low-communication-energy-consumption model deployment system through the collaborative operation of edge nodes and the central training platform. During system operation, edge nodes first continuously collect multi-dimensional data from the operating site, including physical quantities such as temperature, humidity, light intensity, noise, current, voltage, and power provided by environmental sensors; switch status, gear position, executed commands, and feedback quantities reported by the device's internal controller; current bandwidth, latency, packet loss rate, and throughput provided by the network interface; real-time energy consumption and electricity price information output by the local metering module; and computing power status information such as CPU utilization, operating temperature, memory usage, and battery level of the edge system itself. Before entering the algorithm flow, all collected data first undergoes filtering, interpolation, and timestamp calibration by the edge-side preprocessing module, ensuring that it can be input into the index calculation unit at the same time scale after noise suppression. The preprocessed data is then stored in a lightweight time-series buffer for fast access during real-time index calculation and summary generation.

[0078] After the data stabilizes, edge nodes calculate model performance degradation indicators based on real-time inference output and device feedback. By quantifying the magnitude, growth rate, duration, and fluctuation of the error between the predicted results and actual observations, the degree of decline in model adaptability is obtained. Simultaneously, nodes calculate energy consumption deviation based on device operating status and real-time power sequences from energy consumption sensors, combined with historical energy consumption baseline models, thus obtaining energy consumption trend indicators. The system also constructs communication cost indicators based on the current network status and estimated upload costs. Through bandwidth, latency, packet loss rate, and predicted upload volume, it assesses the time, energy, and number of retransmissions that may be consumed in a single upload. These three indicators, combined with preset weights, generate a comprehensive utility value to reflect the system's comprehensive judgment criterion of "the necessity and benefits of model synchronization at this moment." The comprehensive utility value is written to the local historical buffer of the edge nodes over time, serving as input for the automatic threshold adjustment process.

[0079] To enable the judgment criteria to automatically adjust according to node operating status and environmental cycles, edge nodes calculate the historical mean and fluctuation range based on the stored historical comprehensive utility value sequence, and construct an adaptive threshold based on these two factors. This threshold reflects the typical utility range of a node under long-term stable operation, and its value is dynamically updated with load changes, seasonal fluctuations, data drift, and network cycle changes. Nodes use exponential moving averages or fluctuation statistics, allowing the threshold to tighten during stable periods and loosen during volatile periods, thus ensuring that synchronization triggering is neither too frequent nor too delayed. After a new comprehensive utility value is calculated, the node compares it with the latest adaptive threshold. If the comprehensive utility value does not exceed the threshold, the node remains on observation; if the comprehensive utility value exceeds the threshold, the system calculates the difference between the two and generates a synchronization probability based on the magnitude of the difference through a continuous monotonic mapping, with higher synchronization probabilities for greater deviations. The node then generates a pseudo-random number and compares it with the synchronization probability to determine whether to enter the upload process. This mechanism effectively avoids bandwidth congestion caused by multiple nodes simultaneously uploading data during minor fluctuations, while triggering synchronization with a high probability during significant anomalies. In situations involving safety or critical control, such as severe sensor distortion, equipment self-test failure, or temperature exceeding limits, nodes will enter an emergency trigger channel, uploading data directly without probability judgment to ensure the center is aware of critical scenarios immediately. Once synchronously triggered, edge nodes do not upload raw data; instead, they select high-value information from a short-term buffer to construct a priority summary package. The buffer records recent feature inputs, model inference outputs, actual feedback, error statistics, inference latency, inference energy consumption, and confidence levels. Nodes score all candidate samples based on sample error, feature novelty, model uncertainty, and communication cost, selecting the few highest-scoring samples as summary objects. Nodes construct a compact description for each selected sample, including quantified representations of key features, model predictions and actual feedback, error values, timestamps, and a set of sparse gradient signatures to represent local model change trends, enabling the center to determine model degradation locations. Nodes also generate statistical data, such as model confidence distribution, energy consumption change distribution, and inference latency distribution, as well as event index markers for abnormal periods, allowing the center to understand the overall operational status of the nodes without requiring raw data. All summary content is compressed, serialized, and encrypted to form a structured, small-volume summary package.

[0080] If the anomaly is not an emergency, the digest packet will enter the opportunity window queue to await a low-cost upload opportunity. Edge nodes continuously predict future bandwidth, electricity prices, and communication load, forming a sequence of upload costs for future time slices based on the prediction results, and selecting the time period with the lowest cost as the opportunity window. When entering the opportunity window, nodes upload several digest packets sequentially according to digest priority. If bandwidth is limited, only high-priority digests are sent; if the predicted bandwidth increases, nodes can upload in batches to fully utilize the low-cost window. During the upload process, nodes can transmit digests in chunks to reduce retransmission overhead, which are then reassembled and verified by the center. After receiving digest packets from different edge nodes, the center writes them into the data lake and performs structured parsing. Based on the error growth, energy consumption anomalies, bandwidth consumption, and sparse gradient characteristics in the digests, combined with the node's historical operation records and model version information, the center calculates the node utility value for each node. The center then combines the bandwidth budget, energy consumption budget, and center computing resource budget, using integer programming or utility ranking algorithms to determine the set of nodes that need to update the model, maximizing overall benefits with limited resources.

[0081] The center generates a suitable model version for each selected node based on its hardware information, including full-precision models, pruned models, quantized models, distillation models, or accelerator-optimized models. It then performs differential calculations on the old and new models to generate differential update packets, significantly reducing the size of the distributed data. These differential packets are compressed, signed, and distributed to edge nodes via a secure channel. Upon receiving the packet, the edge nodes verify the signature and checksum, generate a new model through differential synthesis, and perform self-testing inference. If successful, the old model is replaced; otherwise, the model is rolled back and reported to the center. Through this process, the system maintains optimal model energy efficiency while ensuring inference continuity. Ultimately, the entire system achieves optimal energy consumption, minimized communication, and long-term stability in intelligent model deployment through a closed-loop process of "edge-side adaptive perception—summary compression and uploading—center-side optimization and training—differential model distribution—edge-side lightweight update." Specific Implementation Example 4:

[0083] like Figures 1 to 3 As shown, based on the content of the above specific embodiments, the following content is further disclosed:

[0084] To further verify the feasibility of this application in practical use, the method of this application is applied to a real-world case, demonstrating the feasibility and core technical effects of a center-training-edge-inference energy optimization model deployment framework in practical use:

[0085] Application Case 1: Actual Operation Case of Energy-Saving Intelligent Air Conditioning Group Control System:

[0086] In a large office building, 36 smart air conditioners are deployed, each equipped with an edge node (low-power AI module), while training servers are deployed in the central computer room. The air conditioners operate stably under normal circumstances, but energy consumption and model degradation frequently occur during seasonal changes.

[0087] After the system is put into operation, the edge nodes begin to collect data on room temperature, humidity, compressor current, fan speed, real-time power, electricity price, network bandwidth, and inference output error. The preprocessing module then filters, calibrates, and aligns all the data in time.

[0088] Within two days of the seasonal change, outdoor temperatures dropped significantly. Air conditioning sampling data showed that the model predicted a target room temperature of 24℃, while the actual measured temperature was 23℃, with the error gradually increasing, the mean error rising from 0.6℃ to 1.9℃. The model degradation index calculated on the edge side showed a significant increase. Meanwhile, the energy consumption baseline model showed an increase in compressor operating cycles, with hourly power consumption 12% higher than the seasonal baseline, and energy consumption change indicators also increased significantly. However, network bandwidth was relatively strained during peak office hours, and communication cost indicators remained at a moderate level. After considering the three indicators, the comprehensive utility value of the edge node exceeded the adaptive threshold for the first time, but the exceedance was small. After a probability-triggered mechanism, the value was not uploaded but only recorded in the historical sequence for continued observation.

[0089] On the morning of the third day, the increased temperature difference and frequent entry into the conference room caused the model error to rise sharply to 2.8℃–3.4℃ within 30 minutes. Simultaneously, the compressor load further increased, with energy consumption increasing by 17%. The difference between the comprehensive utility value and the threshold increased significantly, and the synchronization probability reached 0.82 after probability mapping. The random number was 0.31, lower than the synchronization probability, triggering the node to upload the summary. The node selected 12 highest-priority samples from the buffer of the past 20 minutes. These samples exhibited the following characteristics: a significant decrease in model output confidence; increased lag in temperature feedback; an abnormally high number of compressor start-stop cycles; and a significant difference in feature distribution compared to the training set. The summary package included: quantified values ​​of key features for the 12 samples, model predictions for each sample, actual feedback and error information, a histogram of energy consumption statistics for the increased compressor load, sparse gradient signatures of newly emerging feature drift, and an index of events in the 15-minute period of rapid error growth. The summary, after compression, was only 32KB and was successfully uploaded through a low-energy window (higher bandwidth during lunch break).

[0090] After receiving the abstract, the center discovered through digital twin simulation that the model was not adaptable enough to the "seasonal switching between cold and warm" data. Utility analysis showed that updating the model of this node was expected to reduce compressor power by 9-12% and reduce the average prediction error by 1.8℃. The center added this node along with 5 other similar nodes to the update set, generated a quantized model version adapted to different chips, and distributed it through differential method, with only a 2.1MB differential package being distributed (the complete model is about 12MB).

[0091] After receiving the differential packets, the edge nodes perform verification and synthesis, run self-testing inference in an isolated environment, and switch to the new model after passing the test. Subsequently, the prediction error of the nodes stabilized at around 0.7℃, and the energy consumption decreased by about 11% compared to the previous week.

[0092] Application Case 2: Actual Operation Case of Energy-Saving Intelligent Air Conditioning Group Control System:

[0093] A factory has 24 high-power exhaust fans deployed in its industrial plant. Each fan is equipped with an edge node responsible for energy consumption prediction and airflow regulation model inference to reduce energy waste caused by excessive fan operation at night. The central server is responsible for model training and unified strategy distribution. During the high temperatures of summer, the humidity, dust concentration, and worker activity intensity inside the plant increase, leading to instability in the fan regulation strategy. The edge nodes, while collecting data, found that the deviation between the target airflow predicted by the fan model and the actual airflow measured by the sensors increased daily, the model output confidence level decreased significantly, the fan operating power was 15-22% higher than historical levels for the same temperature period, network bandwidth increased at night, and bandwidth was extremely low during the day due to PLC bus interference. The edge nodes calculated performance degradation indicators, energy consumption change indicators, and communication cost indicators to obtain a comprehensive utility value. Due to poor network conditions and high communication costs during the day, a threshold adaptive mechanism automatically raised the trigger threshold to avoid frequent uploads during the day.

[0094] At 11 PM, the bandwidth prediction module determined that the link bandwidth had significantly increased and the communication cost had significantly decreased. At this time, the comprehensive utility value had exceeded the threshold for approximately 6 consecutive hours, with a large difference. After probability mapping, the synchronization probability reached 0.95, and the random number was 0.21, meeting the upload conditions. The node selected 20 of the most representative samples from the buffer and extracted a compact summary of the high-dimensional wind turbine features, including: real-time wind pressure, air volume, wind turbine speed and corresponding model prediction errors, feature drift reflecting changes in wind resistance caused by dust accumulation on the blades, peak and rate of change statistics of the wind turbine's instantaneous power, a 50-minute event index of the continuous increase in error caused by a sudden increase in wind resistance, and sparse gradient signatures of local drift in the model's internal weights (only recording 27 change dimensions). The compressed summary package was only 45KB and was successfully uploaded at low cost during the night through the opportunity window.

[0095] After analyzing the abstract, the center found that the model did not include the feature of "dust accumulation on blades → increased wind resistance".

[0096] In high-humidity environments, the wind turbine thrust curves differ significantly from the training set. Motor aging causes the power curves to deviate from the standard model. Updating the model can reduce wind turbine energy consumption by 13-18%. The central cluster selects 10 out of 24 wind turbines as priority update targets based on utility ranking and generates corresponding quantized optimization models. Using a differential method, only 1.3MB of data is needed to complete the model update. After hot-updating the model at the edge nodes: the wind turbine airflow error decreased from 12% to 3%, the nighttime wind turbine operating time decreased by 20%, the average energy consumption per wind turbine decreased by 15.6%, and feature shifts caused by dust accumulation can be automatically detected without manual inspection. This case demonstrates the system's adaptability to complex factors such as equipment aging, feature drift, abnormal energy consumption, bandwidth limitations, and opportunistic window optimization in real-world industrial scenarios.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A center training-edge inference energy-optimization model deployment framework, characterized in that: The framework includes the following steps: Sp1: Edge-side index calculation: The operational data collected by the edge nodes are used to calculate the first index representing the degree of model performance degradation, the second index representing the trend of energy consumption change, and the third index representing the communication cost. The three indexes are combined according to preset weights to obtain the comprehensive utility value. Sp2: Adaptive threshold generation: Calculates the historical mean and fluctuation range based on the fluctuation of historical comprehensive utility values, and generates the adaptive threshold for the current node accordingly; Sp3: Probability Trigger Judgment: When the comprehensive utility value is higher than the adaptive threshold, the synchronization probability is calculated based on the difference between the two, and the upload of summary information is determined by comparing the synchronization probability with the random number to avoid concentrated uploads caused by short-term fluctuations. Sp4: Priority Summary Generation: The collected data are prioritized based on factors such as anomaly degree, novelty degree, model output confidence and communication overhead. Several high-priority data are selected to form a summary package. The summary package includes sample feature statistics, model parameter change summary or sparse gradient change identifier. Sp5: Opportunity Window Scheduling: When the emergency upload conditions are not met, multiple low-cost communication periods are identified based on network bandwidth prediction, electricity price prediction, and communication load prediction, and a summary packet is transmitted during the communication periods. Sp6: Central side scheduling decision: The central side receives summary packets from multiple edge nodes, calculates the node utility value based on the model degradation information, energy consumption change information and communication consumption information contained therein, and determines the set of edge nodes that need to be updated under the central resource budget constraints. SP7: Differential Model Distribution: The central side sends the differential update content of the training model to the corresponding edge nodes according to the scheduling results, and generates multiple versions of quantized models in combination with the hardware type of the edge nodes to improve the energy consumption efficiency and execution efficiency of the inference stage.

2. The energy optimization model deployment framework of center training-edge inference according to claim 1, characterized in that: The edge-side index calculation is based on the output error change of the continuous time window data calculation model, the sampling change of device energy consumption, and the estimated value of communication overhead, to generate the first index, the second index, and the third index.

3. The energy optimization model deployment framework based on center training and edge inference according to claim 1, characterized in that: The adaptive threshold generation uses an exponential moving average or an adaptive update rule based on historical utility value fluctuations to generate an adaptive threshold, so that the triggering conditions automatically converge or relax as the node's operating status changes.

4. The energy optimization model deployment framework based on center training and edge inference according to claim 1, characterized in that: The probability triggering determination is based on the magnitude by which the comprehensive utility value exceeds the adaptive threshold to construct a monotonically increasing probability, so that the higher the comprehensive utility value, the greater the corresponding synchronization probability.

5. The energy optimization model deployment framework based on center training and edge inference according to claim 1, characterized in that: The priority summary generation performs scoring based on error value, feature change degree, model output confidence and upload bandwidth cost, and selects data to form a summary package within the target size limit.

6. The energy optimization model deployment framework based on center training and edge inference according to claim 1, characterized in that: The summary package contains at least one of the following: statistical features of high-loss samples; sparse differencing of model weight changes; compact statistical structure of feature anomaly degree; and event index structure of abnormal periods.

7. The energy optimization model deployment framework of center training-edge inference according to claim 1, characterized in that: The opportunity window scheduling determines multiple potential low-cost transmission periods based on future bandwidth prediction, electricity price prediction, or communication energy consumption prediction, and performs non-urgent digest packet transmission within the specified periods.

8. The energy optimization model deployment framework based on center training and edge inference according to claim 1, characterized in that: The central-side scheduling decision selects update nodes that meet energy consumption and bandwidth constraints by sorting node utility values ​​or using integer programming under resource budget constraints.

9. The energy optimization model deployment framework of center training-edge inference according to claim 1, characterized in that: The differential model distributes model update content in a differential manner and generates multiple versions of quantized models based on the hardware type of the edge nodes to reduce energy consumption during the model update and inference processes.

Images