Method and system for controlling power consumption of edge computing unit system

Abstract

The embodiment of the invention provides a power consumption control method and system for an edge computing unit system. The method is applied to the technical field of edge computing, and comprises the following steps: collecting historical input data of edge computing equipment, establishing an LSTM time sequence prediction model at a cloud end, and deploying the LSTM time sequence prediction model on the edge computing equipment; integrating the prediction model into edge device firmware, deploying a lightweight random forest classifier to determine task output, dividing a device operation time period, and completing device operation state setting; constructing an equipment adjustment operation mechanism, performing operation adjustment on the edge equipment, and collecting feedback data; and periodically updating parameters of the LSTM time sequence task prediction model by adopting an incremental learning method based on the feedback data. In this way, accurate task prediction, intelligent scheduling decision and low-overhead model deployment can be achieved on the edge side, and balance is achieved between prolonging of the service life of equipment and guaranteeing of task service quality by reducing the number of invalid wake-up events.

Description

Technical Field

[0001] The present disclosure relates to the field of edge computing technology, and in particular to a method and system for controlling power consumption of an edge computing unit system. Background Art

[0002] In the field of Internet of Things (IoT) systems and edge computing, devices are often deployed in resource-constrained environments and rely on battery power or low-power designs for long-term operation. In order to extend the life cycle of devices, traditional power management strategies generally adopt the "idle sleep" mechanism, that is, when the device has no tasks to process, it enters a low-power sleep state by turning off or reducing the power supply voltage of the main chip, thereby reducing static power consumption. This method reduces the energy consumption baseline of the device to a certain extent and becomes the basic idea for extending the life of the IoT system.

[0003] However, the existing technology has significant defects: frequent sleep-wake-up switching will introduce additional energy consumption overhead. Specifically, when the device wakes up from a sleep state, it needs to go through processes such as voltage recovery, clock stabilization, and peripheral initialization. The transient power consumption at this stage may be much higher than the steady-state operating power consumption. In scenarios with intensive or highly random task requests (such as intelligent monitoring and industrial sensor networks), devices may repeatedly switch states due to frequent responses to sudden tasks, resulting in increased energy consumption. Studies have shown that for some low-power chips, the energy consumption of a single wake-up process is equivalent to the power consumption of continuous operation for several seconds to tens of seconds. If the number of wake-ups per hour exceeds the threshold, the overall energy consumption is even higher than that of a continuously active state.

[0004] Existing solutions attempt to optimize wake-up strategies through static sleep cycle adjustment or simple threshold judgment, but they still have the following limitations: Static scheduling is not adaptable: fixed sleep intervals cannot dynamically respond to changes in task load, resulting in excessive wake-ups at low loads or response delays at high loads. Task processing is fragmented: scattered tasks trigger independent wake-up events, and there is a lack of task merging and batch processing mechanisms, resulting in repeated energy consumption. Insufficient prediction capabilities: traditional heuristic algorithms (such as the moving average method) are difficult to accurately predict non-periodic or sudden tasks, resulting in invalid wake-ups.

[0005] In recent years, although some studies have attempted to introduce machine learning models to predict task arrival patterns, most solutions rely on centralized training in the cloud, do not fully consider the resource constraints of edge devices (such as memory and computing power limitations), and the accuracy loss after model quantization deployment has not been effectively compensated. In addition, existing technologies do not pay enough attention to the coordinated optimization of task classification and priority scheduling, and cannot ensure the real-time performance of high-priority tasks while reducing the number of wake-up times.

[0006] Therefore, there is an urgent need for a power consumption control method that can achieve accurate task prediction, intelligent scheduling decisions, and low-overhead model deployment on the edge side. Summary of the invention

[0007] The present disclosure provides a method and system for controlling power consumption of an edge computing unit system, which achieves a balance between extending the life of the device and ensuring the quality of mission services by reducing the number of invalid wake-up events, and at least solves the technical problems of the existing methods such as high wake-up difficulty, insufficient coordination, and insufficient prediction ability.

[0008] According to a first aspect of the present disclosure, a method for controlling power consumption of an edge computing unit system is provided, comprising the following steps: Collect historical input data of edge computing devices, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing devices; Integrate the constructed prediction model into the edge device firmware, deploy a lightweight random forest classifier to determine the task output, divide the device operation time period, and complete the device operation status setting; Build an equipment adjustment operation mechanism based on the set equipment operation status, adjust the operation of edge equipment, and collect feedback data; Based on feedback data, the parameters of the LSTM time series task prediction model are periodically updated using an incremental learning method.

[0009] According to the aspects described above and any possible implementation method, an implementation method is further provided, wherein the historical input data of the edge computing device includes: task arrival time, task data volume, and power consumption data.

[0010] According to the above aspects and any possible implementation, an implementation is further provided, wherein the LSTM time series task prediction model includes an input layer, an LSTM layer, and an output layer; The input layer inputs historical input data of the edge computing device, the LSTM layer is used to capture dynamic changes in the time series, and the output layer outputs the task prediction situation for the next time step.

[0011] According to the above aspects and any possible implementation, an implementation is further provided, wherein the input layer performs z-score normalization on a three-dimensional tensor consisting of the number of samples, the time step, and the number of features, and generates a continuous time series sliding window in chronological order; The LSTM layer includes a first LSTM layer and a second LSTM layer stacked on each other, the first LSTM layer is used to process the original input sequence and extract local time dependencies, and the second LSTM layer is used to further capture long-term dependencies and global patterns based on the output results of the first LSTM layer; The output layer adopts a linear activation function, and the output result of the output layer is the predicted task arrival time and density in the future time window.

[0012] For the aspects and any possible implementation described above, a further implementation is provided. The LSTM time series task prediction model includes a loss function, and the construction process of the loss function is as follows: Based on the actual task arrival time and the task arrival time predicted by the prediction model, a mean squared error measurement is used to calculate the time prediction error loss function; Based on the one-hot encoding of the actual density category and the probability distribution of the density category predicted by the prediction model, a cross-entropy loss method is used to calculate the task density loss function; The time prediction loss function and the task density loss function are weighted and combined to obtain a joint objective function; Construct an L2 regularization to constrain the model weights and combine it with the joint objective function to obtain the loss function.

[0013] For the aspects and any possible implementation described above, a further implementation is provided. The task output includes simple tasks and complex tasks. The simple tasks are low-load computing and high-tolerance delay tasks, and the complex tasks are high-computing load and low-latency requirement tasks.

[0014] For the aspects and any possible implementation described above, a further implementation is provided. The division of the device operation period is based on the task arrival time and the density output by the LSTM time series task prediction model. The specific process is as follows: When the predicted task density Dt > Dth in a certain period, it is determined as the wake-up period. When the predicted task density Dt < Dth in a certain period, it is the sleep period, where Dth is the threshold.

[0015] For the aspects and any possible implementation described above, a further implementation is provided. The process of constructing a device adjustment operation mechanism based on the set device operation state and adjusting the operation of the edge device is as follows: During the operation of complex tasks and the wake-up period, combined with the predicted task density of the prediction model and detecting the load situation of the model, a judgment is made on whether to save power. If the set voltage-frequency pair is reached, the voltage is adjusted and the clock generator is synchronized to set the frequency, and the switching order of the voltage and frequency is ensured; During the period of no task operation, simple task operation, and sleep stage, monitor and process simple tasks, and set a complex task cache queue and complex task trigger conditions.

[0016] For the aspects and any possible implementation described above, a further implementation is provided. The process of periodically updating the parameters of the LSTM time series task prediction model based on feedback data and using an incremental learning method is as follows: Calculating a model gradient based on the local data samples and labels in the feedback data, and clipping the gradient to obtain a clipped gradient; Gaussian noise is added to the clipped gradient, and multi-device gradient aggregation is performed, and global model parameters are updated based on the aggregated gradient.

[0017] According to a second aspect of the present disclosure, there is provided an edge computing unit system power consumption control system, comprising: Data collection module, task classification module, dynamic scheduling module and model updating module; The data acquisition module is used to collect historical input data of the edge computing device, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing device; The task classification module is used to integrate the constructed prediction model into the edge device firmware, deploy a lightweight random forest classifier to determine the task output, and divide the device operation time period to complete the device operation status setting; The dynamic scheduling module is used to build a device adjustment operation mechanism based on the set device operation status, adjust the operation of the edge device, and collect feedback data; The model updating module is used to periodically update the parameters of the LSTM time series task prediction model based on feedback data and using an incremental learning method.

[0018] Compared with the prior art, the present invention has the following technical effects: The present invention achieves a balance between extending device life and ensuring task service quality through data-driven intelligent scheduling by reducing the number of invalid wake-up events. In addition, the present invention aims to reduce the number of invalid device wake-ups through data-driven intelligent scheduling, thereby improving energy efficiency, extending device life, and meeting the power consumption management requirements of edge computing devices.

[0019] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which: Figure 1A schematic diagram of a method for controlling power consumption of an edge computing unit system according to an embodiment of the present disclosure is shown; Figure 2 A schematic diagram of the structure of a power consumption control system of an edge computing unit system according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Reference Figure 1 As shown, this embodiment provides a method for controlling power consumption of an edge computing unit system, comprising the following steps: S101. Collect historical input data of edge computing devices, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing device.

[0024] In this embodiment, the historical input data of the edge device received by the cloud is used as a data set for LSTM model training. The specific historical data includes: 1. Task arrival time, recording the generation timestamp of each task. 2. Task data volume, the data size of each task. 3. Power consumption data, real-time power consumption during task processing. The data set is divided into training set and test set in chronological order and in a ratio of 7:3 to avoid future data leakage to the training set. The training set is used for model training, and the test set is used to evaluate the difference between the model prediction value and the true value to verify its prediction performance and accuracy.

[0025] Specifically, the LSTM time series task prediction model includes: input layer, LSTM layer, and output layer. The input layer inputs the historical task data of the device, the LSTM layer is used to capture the dynamic changes in the time series, and the output layer outputs the task prediction situation for the next time step.

[0026] In the input layer, the three-dimensional tensor consisting of the number of samples, time steps, and number of features is preprocessed, and each feature is z-score standardized to generate a continuous time series sliding window in chronological order. The z-score standardization formula is: (1) Among them, x is the original data value, μ is the mean of the data set, and σ is the standard deviation of the data set.

[0027] Specifically, in this embodiment, the mean μ and the standard deviation σ are respectively: (2) (3) Where n is the number of samples in the data set and xi is the value of the i-th sample.

[0028] The LSTM layer is composed of two layers of LSTM stacked together, each with 128 hidden units. Two independent LSTM layers, the output of the previous layer is used as the input of the next layer. Each LSTM layer has its own cell unit, which is used to extract time series features at different levels. The first layer of LSTM: processes the original input sequence and extracts local time dependencies. The second layer of LSTM: further captures long-term dependencies and global patterns.

[0029] The output layer uses a linear activation function, and the output result gives the predicted task arrival time and density in the future time window. Time prediction: The task arrival time points {t1, t2, …, t N Density classification: The task density for the next hour is P=[p1,p2,p3].

[0030] Model training and optimization will affect the final prediction error. Traditional LSTM models may not be able to effectively capture the complex relationship between different features when processing multivariate time series. Features such as task arrival time and density do not fully utilize the correlation between the two. For this reason, a joint optimization strategy is adopted to combine the continuous prediction of task arrival time and the discrete classification of density, while considering the correlation between the two. During the training process, the objective function (loss function) is defined, with minimizing the prediction error as the main optimization goal.

[0031] The objective function of the LSTM time series task prediction model jointly optimizes the mean square error of time prediction and the cross entropy loss of density classification to ensure that the model can accurately predict the task arrival pattern while effectively supporting dynamic power scheduling. In practical applications, α and λ are adjusted through experiments to achieve the best balance between prediction accuracy and generalization ability.

[0032] The loss function of task arrival time prediction uses mean square error (MSE) to measure the error of time prediction: (4) in, is the actual arrival time of the i-th task, is the task arrival time predicted by the model, and N is the number of samples.

[0033] The prediction loss function of the task density is divided into discrete categories (low, medium, and high density), using cross-entropy loss: (5) in, is the one-hot encoding of the true density category, is the density category probability distribution predicted by the model, and C is the number of categories.

[0034] In this embodiment, One-Hot Encoding refers to converting categorical variables into a format suitable for processing by machine learning algorithms. For example, the task density categories here are divided into low, medium, and high, and a binary feature vector is created for each category. The vector length is equal to the number of categories, and for each category, only the corresponding feature position is 1, and the rest of the positions are 0. Example: low density [1, 0, 0], medium density [0, 1, 0], high density [0, 0, 1].

[0035] The joint objective function combines the loss of time prediction and density classification in a weighted manner to balance the optimization objectives of the two: (6) Among them, α is the weight coefficient, which adjusts the importance of time prediction and density classification (α=0.7 means more attention to time accuracy). When α→1, the model focuses on minimizing the time prediction error; when α→0, the model focuses on improving the accuracy of density classification. To prevent overfitting and ensure the efficient operation of the model on edge devices, L2 regularization is added to constrain the model weights: (7) Among them, wk is the model weight parameter and λ is the regularization coefficient.

[0036] The final overall objective function (loss function) expression is: (8) After the model training is completed, quantization compression is performed to reduce the size of the model and the computing resource requirements, making it suitable for running on resource-constrained edge devices. The floating-point model is converted to TFLite format using TensorFlow Lite Converter. The prediction error of the model on the test set before and after quantization is compared to ensure that its MSE increases by no more than 5%. The quantized TFLite model is then integrated into the edge device firmware.

[0037] S102: Integrate the constructed prediction model into the edge device firmware, deploy a lightweight random forest classifier to determine the task output, divide the device operation time period, and complete the device operation status setting.

[0038] In this embodiment, a lightweight random forest classifier is deployed in the coprocessor (MCU) of the edge computing device. Based on historical power consumption data and task characteristics (including computational complexity, real-time requirements, and data volume), the device functional tasks are classified into two outputs: 0 (simple tasks) and 1 (complex tasks). Simple tasks refer to tasks with low computational load and high tolerance for latency (such as sensor data filtering and heartbeat packet response), which are processed by the low-power coprocessor (MCU). Complex tasks refer to tasks with high computational load and low latency requirements (such as AI inference and video encoding), which are processed by the main chip.

[0039] The TFLite quantization model of the LSTM time series task prediction model is deployed to the main chip of the edge computing system, and the main chip is responsible for performing computational, storage, and control tasks. Based on the task arrival density statistically obtained from historical data, that is, the number of tasks per minute, a threshold Dth is set. The main chip divides the device operation cycle into a wake-up period and a sleep period by comparing the predicted task arrival time and density of the model. The specific discrimination method is that when the predicted task density Dt > Dth in a certain period, it is the wake-up period. Conversely, when the predicted task density Dt < Dth in a certain period, it is the sleep period. The main chip sends the divided period to the MCU through the PCIe high-speed bus. The wake-up signal is sent by the MCU to the PMIC through the GPIO interface, and the PMIC restores the power supply of the main chip within 10 ms, starts the clock and peripheral devices, and the main chip switches from the deep sleep mode (power consumption < 5 mW) to the active mode (power consumption > 500 mW). The sleep signal is sent by the main chip to the PMIC through I²C to send a sleep instruction and re-enter the deep sleep.

[0040] S103. Construct a device adjustment operation mechanism based on the set device operation state, perform operation adjustment on the edge device, and collect feedback data.

[0041] In this embodiment, the operation strategy is constructed through the divided wake-up period and sleep period, simple tasks, and complex tasks. Specifically: During the period of intensive task prediction, the main chip remains in the active state to efficiently process tasks. During the period of sparse task prediction, the main chip enters the deep sleep mode to reduce power consumption. At the same time, before the predicted task arrives, the system wakes up the main chip in advance to ensure that the device is ready, thereby reducing the response delay and optimizing the overall energy efficiency.

[0042] During the wake-up of the intensive task main chip, by combining the model to predict the task density and real-time monitoring of its own load, a decision is made whether to reduce or increase performance to save power. Specifically, the main chip selects the preset voltage-frequency pair: 1. Low load, voltage 0.8V, frequency locked at 800MHz; 2. Medium load, voltage 1.0V, frequency 1.5GHz; 3. High load, voltage 1.2V, frequency 2.0GHz. The voltage and frequency adjustment instructions are sent to the PMIC through the I²C interface. After receiving the instructions, the PMIC adjusts the output to the target voltage (such as from 1.2V to 0.8V), and synchronously configures the clock generator (PLL) to set the frequency (such as from 2.0GHz to 800MHz). At the same time, the PMIC will ensure the switching order of voltage and frequency (boost first and then increase the frequency, and reduce the frequency first and then reduce the voltage) to prevent chip instability. In the above process, the PMIC cooperates with the main chip to achieve dynamic voltage and frequency adjustment (DVFS), taking into account performance and energy efficiency.

[0043] During the sleep period of the main chip when no tasks are predicted, the low-power coprocessor (MCU) monitors and processes simple tasks in real time; if a complex task arrives, the MCU caches it in the task queue. Only when one of the following wake-up conditions is triggered: the number of queue tasks reaches the preset threshold (50 tasks) or enters the peak period predicted by the model, the MCU sends a wake-up signal to the PMIC, and the PMIC restores power to the main chip within 10ms and starts the clock and peripherals.

[0044] S104. Based on the feedback data, the parameters of the LSTM time series task prediction model are periodically updated using an incremental learning method.

[0045] The cloud periodically updates the LSTM model parameters based on the feedback data, and optimizes the prediction accuracy through incremental learning. The updated model is re-quantized and deployed to the edge device to form a closed-loop optimization system. Traditional incremental learning simply collects data in the cloud, retrains the model, and then pushes updates. This method has data privacy issues. Taking this issue into consideration, the present invention uses edge-side data desensitization during the upload process, first performs differential privacy, and then uploads it to the cloud. In this embodiment, feedback data refers to two types of data, both generated by edge devices. 1. The difference between the predicted results and the actual results of the lightweight model deployed on the local device is used to guide the update of the cloud model parameters; 2. The original time series data such as the arrival time, type, data volume, power consumption, etc. of the newly collected tasks within a period of time is used to expand the training data set.

[0046] The specific implementation method is that the edge device uses local data to calculate the model gradient, clips the gradient to prevent the gradient amplitude from being too large, and limits its L2 norm to not exceed the threshold C. Gaussian noise N(0,σ²) is added to the gradient to satisfy ( ,δ)-differential privacy. Finally, the cloud aggregates the noise gradients of multiple devices and updates the global model parameter w. The privacy budget consumption is further reduced by randomly sampling only some devices to participate in the update in each round. This method ensures the model prediction accuracy while protecting data privacy.

[0047] The relevant calculation formulas are as follows: Gradient calculation formula: (9) in, is the loss function, are model parameters, are local data samples and labels.

[0048] Gradient clipping formula: (10) Among them, C is the threshold.

[0049] The calculation formula of noise scale σ is: (11) Where δ is the failure probability, Budget for privacy. Gradient expression after adding noise: (12) Aggregate formula: (13) Model update formula: (14) in, is the learning rate.

[0050] like Figure 2 As shown, this embodiment also provides an edge computing unit system power consumption control system, including: a data acquisition module 1, a task classification module 2, a dynamic scheduling module 3 and a model updating module 4; The data collection module 1 is used to collect historical input data of the edge computing device, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing device; Task classification module 2 is used to integrate the constructed prediction model into the edge device firmware, deploy a lightweight random forest classifier to determine the task output, and divide the device operation time period to complete the device operation status setting; The dynamic scheduling module 3 is used to build a device adjustment operation mechanism based on the set device operation status, adjust the operation of the edge device, and collect feedback data; The model updating module 4 is used to periodically update the parameters of the LSTM time series task prediction model based on feedback data and using an incremental learning method.

[0051] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present disclosure is not limited by the order of the actions described, because according to the present disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present disclosure.

[0052] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in this application can be achieved, and this document is not limited here.

[0053] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for controlling power consumption of an edge computing unit system, characterized in that: It includes the following steps: Collect historical input data of the edge computing device, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing device; Integrate the constructed prediction model into the edge device firmware, deploy a lightweight random forest classifier to determine the task output, and divide the device operation period to complete the setting of the device operation state; Construct a device adjustment operation mechanism based on the set device operation state, perform operation adjustment on the edge device, and collect feedback data; Periodically update the parameters of the LSTM time series task prediction model based on the feedback data and using the incremental learning method.

2. The edge computing unit system power consumption control method according to claim 1, characterized in that: The historical input data of the edge computing device includes: task arrival time, task data volume, and power consumption data.

3. The edge computing unit system power consumption control method according to claim 1, characterized in that: The LSTM time series task prediction model includes an input layer, an LSTM layer, and an output layer; The input layer inputs the historical input data of the edge computing device, the LSTM layer is used to capture the dynamic changes in the time series, and the output layer outputs the task prediction situation for the next time step.

4. The edge computing unit system power consumption control method according to claim 3, characterized in that: The input layer performs z-score normalization on the three-dimensional tensor composed of the number of samples, time step length, and number of features, and generates a continuous time series sliding window in chronological order; The LSTM layer includes a first LSTM layer and a second LSTM layer stacked on top of each other. The first LSTM layer is used to process the original input sequence and extract local time dependencies, and the second LSTM layer is used to further capture long-term dependencies and global patterns based on the output result of the first LSTM layer; The output layer uses a linear activation function, and the output result of the output layer is the predicted task arrival time and density for the future time window.

5. The edge computing unit system power consumption control method according to claim 4, characterized in that: The LSTM time series task prediction model includes a loss function, and the construction process of the loss function is as follows: Based on the true task arrival time and the task arrival time predicted by the prediction model, use the mean squared error measure to calculate the time prediction error loss function; Based on the one-hot encoding of the true density category and the probability distribution of the density category predicted by the prediction model, use the cross-entropy loss method to calculate the task density loss function; Perform weighted combination on the time prediction loss function and the task density loss function to obtain a joint objective function; Construct an L2 regularization constraint model weight and combine it with the joint objective function to obtain a loss function.

6. The edge computing unit system power consumption control method according to claim 4, characterized in that: The task output includes simple tasks and complex tasks. The simple tasks are low-load computing and high-tolerance delay tasks, and the complex tasks are high-computing load and low-latency requirement tasks.

7. The edge computing unit system power consumption control method according to claim 6, characterized in that: The division of the device operation period is based on the task arrival time and the density output by the LSTM time series task prediction model. The specific process is as follows: When the predicted task density Dt > Dth in a certain period, it is determined as the wake-up period. When the predicted task density Dt < Dth in a certain period, it is the sleep period, where Dth is the threshold.

8. The method for controlling the system power consumption of the edge computing unit according to claim 7, wherein, The process of constructing a device adjustment operation mechanism based on the set device operation status and adjusting the operation of the edge device is as follows: During the operation of complex tasks and the wake-up period, the predicted task density of the prediction model is combined with the detection of the model load to determine whether power consumption is saved. If the set voltage-frequency pair is reached, the voltage is adjusted and the clock generator is synchronously configured to set the frequency, and the switching order of voltage and frequency is ensured; In the no-task-running, simple-task-running and sleep stages, simple tasks are monitored and processed, and complex-task cache queues and complex-task triggering conditions are set.

9. The edge computing unit system power consumption control method according to claim 1, characterized in that: The process of periodically updating the parameters of the LSTM time series task prediction model based on feedback data and using an incremental learning method is as follows: Calculating a model gradient based on the local data samples and labels in the feedback data, and clipping the gradient to obtain a clipped gradient; Gaussian noise is added to the clipped gradient, and multi-device gradient aggregation is performed, and global model parameters are updated based on the aggregated gradient.

10. An edge computing unit system power consumption control system, implemented using the edge computing unit system power consumption control method according to any one of claims 1 to 9, characterized in that: include: Data collection module (1), task classification module (2), dynamic scheduling module (3) and model updating module (4); The data collection module (1) is used to collect historical input data of the edge computing device, establish an LSTM time series prediction model in the cloud, and deploy the LSTM time series prediction model on the edge computing device; The task classification module (2) is used to integrate the constructed prediction model into the edge device firmware, and deploy a lightweight random forest classifier to determine the task output, and divide the device operation time period to complete the device operation status setting; The dynamic scheduling module (3) is used to build a device adjustment operation mechanism based on the set device operation status, adjust the operation of the edge device, and collect feedback data; The model updating module (4) is used to periodically update the parameters of the LSTM time series task prediction model based on feedback data and using an incremental learning method.

Images