Device-oriented energy saving optimization method based on transformer

Abstract

The application provides a device-oriented energy-saving optimization method based on a Transformer, relates to the technical field of energy-saving optimization, and comprises the following steps: S1, establishing a device operation condition model; S2, introducing an energy efficiency model to calculate the energy consumption of the device under different conditions; S3, introducing a multi-objective minimization optimization algorithm, combining a short-term and long-term energy consumption prediction model, and taking into account energy consumption and operation state stability; and S4, introducing a fine-grained real-time parameter adjustment mechanism to adjust the device parameters in real time under different conditions and environmental disturbances; the application combines short-term and long-term energy consumption prediction, combines the energy consumption of the current device state and the energy consumption trend in the future, effectively balances short-term energy saving and long-term operation effect through a multi-objective function, can realize real-time energy saving while ensuring that the overall energy consumption in the future remains at a low level, and avoids the adverse effects of a local optimal solution on global energy consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy-saving optimization, in particular to a device-oriented energy-saving optimization method based on a Transformer. BACKGROUND

[0002] With the rapid growth of global energy consumption, energy-saving optimization has become one of the core tasks in industrial equipment and household equipment control. In recent years, in the face of rising energy costs and environmental protection pressures, the demand for energy efficiency optimization in various industries has increased dramatically. However, due to the variety of devices and complex operating environments, traditional energy-saving methods often fail to meet the energy-saving needs of various devices under different operating conditions. In the prior art, energy-saving optimization methods are usually limited to energy consumption control of a single device or a single target, and lack consideration of complex coupling relationships. For example, many optimization algorithms only focus on minimizing power consumption, without considering other key targets such as running stability and comfort, resulting in optimization results that are difficult to meet the diverse needs of actual applications.

[0003] In addition, devices are dynamically affected by environmental factors (such as temperature and humidity, load changes, etc.) and internal states (such as power and speed, etc.) during operation, resulting in significant fluctuations in energy efficiency curves. This fluctuation directly affects the stability of device operation and the comfort experience of users, and also significantly increases the complexity of energy consumption management. In multi-device coordination or source-grid-load integrated application scenarios, the operating parameters of different devices often have strong coupling relationships, further increasing the difficulty of energy-saving optimization. Existing energy-saving control methods are difficult to achieve coordinated optimization among multiple devices, and thus cannot effectively reduce overall energy consumption while ensuring device stability. Therefore, the present application proposes a device-oriented energy-saving optimization method based on a Transformer to solve the problems in the prior art. SUMMARY

[0004] To solve the above problems, the present application proposes a device-oriented energy-saving optimization method based on a Transformer. This method uses a machine learning model to achieve long-term and short-term prediction of device energy consumption, and combines a multi-objective optimization algorithm to accurately regulate device operating conditions. By considering short-term and long-term energy efficiency, it balances multiple requirements such as energy saving and stability, and can adapt to multi-device and complex environmental multi-objective energy-saving requirements. Through real-time adjustment of device fine-grained parameters, the device can achieve optimal operating conditions while reducing energy efficiency fluctuations, achieving stable and intelligent energy-saving effects, effectively solving the technical problems of complex operating conditions and inaccurate energy efficiency prediction in the prior art.

[0005] To achieve the purpose of the present application, the present application realizes the following technical scheme: a device-oriented energy-saving optimization method based on a Transformer, comprising the following steps:

[0006] S1: Establishing a device operating condition model;

[0007] S2: Introducing an energy efficiency model to calculate the energy consumption of the device under different operating conditions;

[0008] S3: Introducing a multi-objective minimization optimization algorithm, combining short-term and long-term energy consumption prediction models, taking into account energy consumption and operating state stability;

[0009] S4: Introducing a fine-grained real-time parameter adjustment mechanism to adjust device parameters in real time under different operating conditions and environmental disturbances.

[0010] Further improvements are that S1 includes the following steps:

[0011] A Transformer model is used to establish a device operating condition model, which captures the nonlinear relationship and high-order interaction of each parameter in the device operating state by introducing a state vector composed of various operating parameters of the device, i.e. The mathematical description of the device operating state is:

[0012]

[0013] where, is the current state of the device, is the state of the device at the previous time, is the external disturbance (such as temperature, humidity, and other environmental parameters), is a specific parameter of the device, including fixed attributes and adjustable parameters;

[0014] A multi-head attention mechanism is used, with 8 attention heads and a hidden layer dimension of 512. Each Transformer encoder contains a feedforward neural network layer with a dimension of 2048. The Transformer model contains 4 encoder layers to effectively capture the complex interaction between the device state and external disturbances.

[0015] Further improvements are that the training data of the model comes from real device operation log data and simulation platform generated data, including historical operating parameters, environmental variables, and device state records of the device under different operating conditions and environmental conditions. By combining real data and simulation data, the Transformer model can be fully trained to achieve accurate modeling of the device operating state;

[0016] For , a time series formula based on weighted moving average is used to smooth short-term fluctuations while considering historical trends. It is a simple and commonly used method for environmental parameter prediction. The specific formula is as follows:

[0017]

[0018] wherein denotes the historical environment parameter values of the previous time steps, is a weighting coefficient, usually satisfying , is the number of observed historical steps;

[0019] Training through the Transformer model can obtain effective modeling of the device running state, providing data support for further energy efficiency prediction.

[0020] Further improvements are that the S2 comprises the following steps:

[0021] After constructing the device running state model, an energy efficiency model is introduced to calculate the energy consumption of the device under different working conditions, and the energy efficiency score of the device is calculated according to the running state and environmental factors :

[0022]

[0023] wherein is the energy efficiency score based on the current running state of the device, is an influence factor based on the external environment, is a coupling term between the running state and the external factors, and the energy consumption prediction is based on the energy efficiency model, combining short-term and long-term data to predict the energy consumption in the future period of time;

[0024] The short-term energy consumption prediction utilizes historical time series to predict the short-term energy consumption through a recursive manner:

[0025]

[0026] The long-term energy consumption prediction is performed through a multi-step forward Transformer model:

[0027] .

[0028] Further improvements are that the S3 comprises the following steps:

[0029] The multi-objective minimization optimization algorithm takes energy consumption and stability as the core targets, and the objective function is defined as:

[0030]

[0031] wherein, is the short-term energy consumption target, i.e., the total energy consumption in the current prediction period, which is calculated as the cumulative instantaneous energy efficiency model of the device under the current state and environmental parameters,

[0032]

[0033] For long-term energy consumption target, i.e. energy consumption trend prediction in future period, the Transformer model is used for prediction, which represents the total amount of energy consumption of the device in the next period, i.e.

[0034]

[0035] D is the stationarity target, which measures the state fluctuation amplitude of the device during operation, i.e.

[0036]

[0037] and are the weight parameters of short-term energy consumption, stationarity and long-term energy consumption, respectively;

[0038] To solve the multi-objective optimization problem, particle swarm optimization algorithm PSO or genetic algorithm GA is used as the optimization solver;

[0039] In the optimization process, the non-dominated sorting and crowding distance combined Pareto front screening method is used to perform hierarchical screening and reservation on the solution space, so as to obtain the optimal solution set under multi-objective;

[0040] For the above objective function, it is continuously iteratively optimized, i.e. the optimal parameter combination is solved under multi-variable and multi-constraint conditions, so that the device reaches the best balance point of energy saving and stability under the given constraint conditions.

[0041] Further improvement lies in that the particle swarm optimization algorithm PSO or genetic algorithm GA is used as the optimization solver, and the specific parameters are set as follows:

[0042] When PSO is used, the particle swarm size is set to 50, the maximum iteration number is set to 200, the inertia weight is set to 0.7, and the individual learning factor and social learning factor are both set to 1.5;

[0043] When GA is used, the population size is set to 100, the maximum iteration number is set to 300, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the tournament selection strategy is used for individual selection.

[0044] Further improvement lies in that the S4 comprises the following steps:

[0045] By detecting the change of the device state to determine whether adjustment is needed, and performing fine adjustment on the parameters when certain conditions are met, and setting an adjustment threshold For preventing unnecessary frequent adjustment, the parameter adjustment is triggered when the state change exceeds the threshold value, and the specific adjustment rule is as follows:

[0046]

[0047] Wherein represents the parameter adjustment amplitude of the current moment , is an adjustment factor, which determines the size of the adjustment amplitude, is the threshold value of the adjustment.

[0048] Further improvement lies in that the threshold value Specifically, by statistically analyzing the response data of the equipment under stable working conditions, the standard deviation range of the natural fluctuation of the state vector is calculated, and the threshold value is set as the upper limit of the fluctuation range, so as to ensure that the parameter adjustment is triggered only when the system state changes significantly.

[0049] The adjustment factor Specifically, according to the dynamic response characteristics of different equipment, the relationship between the hysteresis time and the response intensity of the system to the parameter adjustment is determined by experiment, a coefficient value balanced between the response efficiency and the energy consumption improvement is selected, and the grid search method or the Bayesian optimization is used for offline parameter adjustment in the interval.

[0050] The beneficial effects of the present application are:

[0051] 1. The present application combines short-term and long-term energy consumption prediction, combines the energy consumption of the current equipment state and the energy consumption trend in the future, balances the short-term energy saving and the long-term operation effect through a multi-objective function, can realize real-time energy saving while ensuring the overall energy consumption in the future to remain at a low level, and avoids the adverse effects of local optimal solution on global energy consumption.

[0052] 2. The present application introduces the stability target D and extends to the long-term prediction step, so that the algorithm can optimize the energy consumption fluctuation and avoid the equipment wear and loss of efficiency caused by frequent parameter adjustment or running state change.

[0053] 3. Through the dynamic parameter adjustment mechanism, the algorithm can finely adjust the equipment parameters according to the real-time environmental changes, so that the running state of the equipment can quickly respond to external disturbances. The fine-grained adjustment mechanism not only prevents over-adjustment, but also realizes more accurate energy saving effect, and further improves the fineness and stability of energy saving optimization.

[0054] 4、The general equipment working condition model constructed based on the Transformer model is suitable for operation scenes of various equipment, has strong adaptability and universality, the model can be flexibly adapted to different equipment types and operation environments, and is widely suitable for various general equipment in a source network load system. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A process flowchart of the present application. DETAILED DESCRIPTION

[0056] In order to deepen the understanding of the present application, the present application will be further described in combination with examples, and the examples are only used to explain the present application and do not constitute a limitation on the protection scope of the present application.

[0057] Example 1

[0058] According to Figure 1 The device-oriented energy-saving optimization method based on Transformer proposed in the present embodiment includes the following steps:

[0059] Device data preprocessing

[0060] The device running state and environmental parameter data obtained from the data acquisition module need to be systematically preprocessed to improve the model input quality and enhance the prediction stability and robustness of the model. The preprocessing process includes the following key steps:

[0061] Abnormal value identification and processing.

[0062] Statistical analysis and rule-based detection methods are used to identify abnormal values.

[0063] Statistical discrimination method: based on the statistical distribution of historical data, if the data at a certain time deviates from the mean value by more than 3 times the standard deviation (3 ) is regarded as abnormal.

[0064] Boundary rule method: set reasonable upper and lower limits combined with the actual working parameter interval of the equipment (such as temperature ∈ [10℃, 45℃]), and determine as false data if it is exceeded.

[0065] For abnormal value processing, linear interpolation or neighborhood mean filling method is used for repair to maintain the continuity of the time series.

[0066] Normalization processing.

[0067] All input variables (especially environmental parameters ) need to be normalized before entering the model to eliminate the dimension difference and improve the convergence speed of the model. The normalization method adopts Z-Score standardization, and the specific formula is as follows:

[0068]

[0069] where, is the mean of the training dataset, is the standard deviation, is the normalized environment parameter value. This method can ensure that the data meets the zero mean and unit variance distribution, which is suitable for the input requirements of various machine learning models.

[0070] Training and prediction of equipment working condition model

[0071] Transformer model training

[0072] The historical running data of the equipment is input into the Transformer model as the training set to train the working condition model. Through model learning, the complex coupling relationship between various parameters of the equipment is obtained, and the energy efficiency performance under different working conditions is obtained.

[0073] Short-term energy consumption prediction

[0074] Using the above trained Transformer model and the time series formula based on weighted moving average, the equipment state at the future time step is predicted and external disturbance parameters , that is

[0075]

[0076]

[0077] The weight can be set according to the influence degree of different time steps. For example, if the recent historical data is expected to have a greater impact, it can be set as is the decreasing coefficient:

[0078]

[0079] Based on this, the short-term energy consumption prediction of the future T steps is carried out, and the calculation formula is as follows, where is the energy consumption prediction value of the equipment under the current working condition.

[0080]

[0081] Long-term energy consumption prediction

[0082] Through the Transformer model, the equipment state and environment parameters at the future step are predicted to obtain the long-term energy consumption target .

[0083]

[0084] Multi-objective optimization process

[0085] Objective function construction

[0086] Combine short-term and long-term energy consumption prediction results to construct a multi-objective optimization objective function:

[0087]

[0088] Where the stationarity objective is defined as the energy consumption fluctuation within the future step:

[0089]

[0090] Optimization solution

[0091] Through genetic algorithm (GA) or particle swarm optimization (PSO), the device operating parameters are iteratively optimized. The optimization process includes individual selection, crossover mutation, fitness evaluation and long-term feedback adjustment, etc. until the objective function converges or reaches the set iteration upper limit.

[0092] Fine-grained real-time parameter adjustment

[0093] Based on the multi-objective optimization result, dynamically adjust the key parameters of the device (such as temperature set point, wind speed, etc.). When the device state changes by more than the threshold , trigger the adjustment mechanism:

[0094]

[0095] The adjusted state will be re-input into the working condition model for energy consumption prediction and optimization, so as to realize real-time energy-saving control of the device in different operating environments.

[0096] Real-time guarantee and deployment scheme

[0097] In order to ensure the real-time and deployable ability of the energy-saving optimization strategy in the actual device running process, the present invention optimizes the algorithm design and system architecture, guarantees the response speed and deployment efficiency, which specifically includes the following two aspects:

[0098] Optimization algorithm response time

[0099] Lightweight Transformer architecture and efficient multi-objective optimizer (such as PSO / GA) are adopted, and the following methods are used to significantly reduce the online calculation delay:

[0100] (1) Model pruning and quantization: parameter pruning and weight quantization are performed on the Transformer model to reduce the computational burden during inference.

[0101] (2) Pre-computation window sliding strategy: For short-term energy efficiency prediction and control parameter adjustment, a sliding window mechanism is used to pre-compute partial results, reducing real-time computation pressure.

[0102] (3) Optimizer iterative control: Limit the number of iterations per round (e.g. ≤ 20 rounds) while ensuring optimization accuracy, and combine heuristic initial solution generation to shorten convergence time.

[0103] According to actual measurement, under typical configuration (Intel i7 CPU, 16GB RAM), the average response time of the complete decision-making cycle of the method is less than 50 ms, which can meet the requirements of most industrial device control systems for real-time scheduling within 100 ms.

[0104] Edge computing deployment architecture

[0105] To achieve high reliability and low latency online control, the method supports deployment on edge computing nodes such as industrial gateways, intelligent edge boxes (Edge Box), etc. The system architecture is as follows:

[0106] (1) Cloud-edge collaborative architecture: Model training and parameter optimization can be performed on the cloud, while prediction and scheduling decisions in the running phase are deployed on edge devices, reducing dependence on the cloud.

[0107] (2) Hardware compatibility: The optimization algorithm is adapted to mainstream embedded devices (such as ARM Cortex-A72 chips) and industrial edge servers (such as NVIDIA Jetson, Raspberry Pi, etc.), ensuring low power consumption and high performance.

[0108] (3) Communication protocol adaptation: Supports standard protocol communication with device-side OPC UA, Modbus, MQTT, etc., achieving seamless integration.

[0109] This architecture not only improves the response speed of algorithmic decision-making and system security, but also enables autonomous operation off the network, suitable for scenarios where the factory network is unstable or edge independent control is required.

[0110] Embodiment Two

[0111] According to Figure 1 , the embodiment proposes a device-oriented energy-saving optimization method based on Transformer:

[0112] To verify the effectiveness of the multi-objective energy-saving optimization method based on Transformer proposed in the present invention, a comparative experiment was designed and carried out. The experiment selected an industrial ventilation system as the object, and under the same environmental disturbance and device task conditions, the following two methods were used for energy-saving control:

[0113] (1) Comparative Method A: LSTM + Single-objective Optimization, using Long Short-Term Memory (LSTM) for energy consumption prediction with the sole optimization goal of minimizing energy consumption;

[0114] (2) Invention Method B: Transformer + Multi-objective Optimization, introducing a Transformer model in energy consumption prediction, combining stationarity objectives with short-term / long-term energy efficiency for multi-objective optimization control.

[0115] Experiments were run for 24 hours on both simulation platforms and real device log datasets with a 5-second sampling interval. The following two key indicators were mainly evaluated:

[0116] (1) Average energy consumption per unit time (kWh).

[0117] (2) State fluctuation index (standard deviation) to measure the stationarity of the running process.

[0118] The experimental results are as follows:

[0119]

[0120] From the experimental results, compared with the traditional LSTM combined with single energy consumption target optimization method, the invention method significantly improves the overall energy saving effect while maintaining device operation stability. Among them, the average energy consumption is reduced by about 20%, and the state fluctuation amplitude is reduced by about 30%, fully verifying the dual optimization capability of the method in energy saving and stationarity.

[0121] To further verify the adaptability of the invention multi-objective energy saving optimization method in complex and dynamic environments, cross-scenario tests were designed, introducing extreme working conditions for simulation, including the following two typical cases:

[0122] (1) Sudden load: the device load suddenly rises to 150% of the rated power in a short time, and then remains at a high load state for 5 minutes;

[0123] (2) Environmental mutation: the environmental temperature suddenly changes from normal temperature (25°C) to high temperature (40°C), and the humidity rises from 40% to 80%, simulating the scenario of rapid deterioration of external climate.

[0124] The test was run in the simulation platform, comparing the performance of the invention method and the traditional LSTM + single-objective optimization strategy under extreme working conditions, mainly evaluating:

[0125] (1) System response time (from disturbance to stable recovery);

[0126] (2) Peak energy consumption (kW);

[0127] (3) state fluctuation intensity (maximum deviation amplitude);

[0128] The experimental results are shown in the following table:

[0129]

[0130] The above tests show that in extreme scenarios such as sudden load and severe environmental changes, the method of the present application can quickly identify changes in working conditions, dynamically adjust parameters, achieve smoother response and faster energy consumption regression, and is significantly better than traditional strategies. This capability is particularly important for industrial equipment, smart homes and other application scenarios that require high stability and comfort.

[0131] The present application combines short-term and long-term energy consumption prediction, combines the current device state energy consumption and future multi-step energy consumption trend, effectively balances short-term energy saving and long-term operation effect through multi-objective function, can realize real-time energy saving while ensuring the overall energy consumption in the future to remain at a low level, and avoids the adverse effects of local optimal solution on global energy consumption. And the present application introduces the stability target D and extends it to the long-term prediction step, so that the algorithm can optimize energy consumption fluctuation and avoid equipment wear and efficiency loss caused by frequent parameter adjustment or running state change. At the same time, through the dynamic parameter adjustment mechanism, the algorithm can make fine adjustments to the device parameters according to the real-time environmental changes, so that the running state of the device can quickly respond to external disturbances, and the fine-grained adjustment mechanism not only prevents over-adjustment, but also achieves more accurate energy saving effect, further improving the fineness and stability of energy saving optimization. In addition, the general device working condition model constructed based on the Transformer model is suitable for the running scene of various devices, has strong adaptability and universality, and the model can flexibly adapt to different device types and running environments, and is widely used in various general devices in the source network load system.

[0132] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A device-oriented energy saving optimization method based on Transformer, characterized in that, The method comprises the following steps: S1: Establishing a device operating condition model using a Transformer model, by introducing device state }, the mathematical description of the device operating state is: ， wherein, is the state of the device at the current time, is the state of the device at the previous time, is the external disturbance, is the device-specific parameters, including fixed attributes and adjustable parameters, the fixed attributes including the adjustable frequency range, rated power, power supply resistance when the variable frequency motor is running; The multi-head attention mechanism is adopted, the number of attention heads is set to 8, the dimension of the hidden layer is set to 512, each Transformer encoder layer comprises a feedforward neural network layer with a dimension of 2048, and the Transformer model comprises 4 encoder layers; S2: After constructing the device operating condition model, introduce the energy efficiency model to calculate the energy consumption of the device under different conditions, according to the operating state and environmental factors of the device to calculate the energy efficiency score of the device: ， wherein is an energy efficiency score based on the current running state of the device, is an influence factor based on external environment, is a coupling term between the running state and the external factor, the energy consumption prediction is based on the energy efficiency model, combining short-term and long-term data to predict the energy consumption in the future period of time; The short-term energy consumption prediction utilizes historical time series to predict short-term energy consumption in a recursive manner: ， The long-term energy consumption prediction is performed through a multi-step forward Transformer model: ， wherein, to calculate the energy efficiency of the device at a certain time t; S3: A multi-objective minimization optimization algorithm is introduced, the short-term and long-term energy consumption prediction models are combined, and a target function is defined as: ， ， ， D is a stationarity target, which measures the state fluctuation amplitude of the equipment during operation: ， and are weight parameters for short-term energy consumption, stationarity, and long-term energy consumption, respectively; To solve the multi-objective optimization problem, a particle swarm optimization algorithm PSO or a genetic algorithm GA is used as an optimization solver; In the optimization process, a Pareto front screening method combining non-dominated sorting and crowding distance is used to perform hierarchical screening and reservation on the solution space, so as to obtain an optimal solution set under multi-objective conditions; For the above target function, iterative optimization is continuously performed, so that the optimal parameter combination can be solved under multi-variable and multi-constraint conditions, and the equipment can reach the best balance point of energy saving and stability under the given constraint conditions; S4: Introduce fine-grained real-time parameter adjustment mechanism, adjust device parameters in real time under different working conditions and environmental disturbances, judge whether adjustment is needed by detecting the change of device state , set adjustment threshold , trigger parameter adjustment when the state change exceeds the threshold, the specific adjustment rules are as follows: ， wherein represents the parameter adjustment amplitude at the current time point , is an adjustment factor, is an adjusted threshold value.

2. The Transformer-based device-oriented energy saving optimization method according to claim 1, wherein: For , a time series formula based on weighted moving average is adopted to smooth short-term fluctuations, and the specific formula is as follows: ， wherein represents the preceding historical environment parameter values at the previous time step, is a weighting coefficient satisfying , is the number of observed historical steps; The Transformer model is trained.

3. The Transformer-based device-oriented energy saving optimization method of claim 1, wherein: The particle swarm optimization algorithm PSO or the genetic algorithm GA is used as an optimization solver, and the specific parameters are set as follows: When the PSO is used, the particle swarm size is set to 50, the maximum iteration number is set to 200, the inertia weight is set to 0.7, and the individual learning factor and the social learning factor are both set to 1.5; When the GA is used, the population size is set to 100, the maximum iteration number is set to 300, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the tournament selection strategy is used for individual selection.

4. The Transformer-based device-oriented energy saving optimization method of claim 1, wherein: Threshold value Specifically, by statistically processing the response data of the equipment under stable working conditions, the standard deviation range of the natural fluctuation of the state vector is calculated, and the upper limit of the fluctuation range is set as the threshold value. for the upper limit of the fluctuation range. Adjusting factor Specifically, according to the dynamic response characteristics of different devices, the relationship between the hysteresis time and the response strength of the system parameter adjustment is determined by experiment, and the coefficient value balanced between the response efficiency and the energy consumption improvement is selected, so that the system parameter adjustment is carried out in the range of The offline parameter adjustment is carried out by the grid search method or the Bayesian optimization in the range.

Images