Deployment and optimization method of net load prediction model on smart energy unit

Abstract

The invention relates to a method for deploying and optimizing a net load prediction model on a smart energy unit, and the method comprises the steps: carrying out the compatible adaption of the smart energy unit and a model, the method specifically comprises the following steps: according to the complexity, data volume and performance requirements of a model and budget research model deployment software and hardware requirements, evaluating required example support, memory and storage space requirements by analyzing the complexity of the model, and according to the model and a framework, determining whether additional dependence files are needed and researching security requirements; carrying out target detection by adopting an ensemble learning algorithm; and a time sequence event model of a self-attention mechanism is introduced to carry out fault early warning and defect elimination guidance, and the potential fault occurrence probability of the key equipment is predicted. The method has the beneficial effects that 1, the suitability of the model and the smart energy unit is improved; and 2, the model performance is improved. And 3, effective model deployment and test are realized. 4, the power market operation and regulation capability can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of load forecasting, and in particular relates to a method for deploying and optimizing a net load forecasting model on a smart energy unit. Background Art

[0002] With the development of smart energy systems, net load forecasting plays a vital role in the operation and regulation of the power market. Neural network models have been widely used in the field of net load forecasting, but such models are highly complex and resource-intensive. The resources of smart energy units are relatively limited, and there are many challenges in deploying net load forecasting models on them, such as the compatibility of the model with the smart energy unit, and the operating efficiency of the model under limited resources. At present, there is a lack of a systematic method to solve the effective deployment and performance optimization of net load forecasting models on smart energy units to meet the needs of various demand response resources on the user side to actively participate in the operation and regulation of the power market. Summary of the invention

[0003] In order to solve the technical problems in the background technology, the present invention proposes a method for deploying and optimizing a net load prediction model on a smart energy unit; the hardware and software requirements for the compatible adaptation model deployment between the smart energy unit and the model: the hardware and software requirements for model deployment are comprehensively considered based on multiple factors such as model complexity, data volume, performance requirements, and budget. By analyzing the model complexity, the required example support, memory, and storage space requirements are evaluated to ensure the normal operation of the process after the model is deployed. Determine whether additional dependency files are required based on the model and framework used, and learn the security requirements for model deployment, such as isolation access control.

[0004] Smart energy unit resource matching evaluation: In-depth study of the operating system and hardware of the smart energy unit, focusing on whether the NPU's support for model examples can meet load forecasting requirements.

[0005] Adaptation method: Since the operating system is the basis for model deployment, corresponding driver support is required for specific hardware requirements, and special dependency files need to match the operating system. Therefore, the adaptation method of the prediction model and the smart energy unit is used to implement the model's call to software and hardware resources. The specific operation is to evaluate the operating system and hardware resources of the smart energy unit to determine whether they meet the requirements, then perform necessary configuration and driver development on the operating system, and deploy dependency files on the model to meet the model requirements.

[0006] Model compression and optimization based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit Model compression method: The model compression method reduces the accuracy loss of the model by exploring the impact of different quantization strategies (such as dynamic quantization, mixed precision quantization and other algorithms) on model performance and storage, determines the quantization compression strategy, and compresses the model in combination with the NPU unit on the heterogeneous computing chip.

[0007] Model optimization method: Minimize the selected metric layer by layer in the hyperspace composed of distribution and feasible maximum and minimum distribution to obtain the optimal quantization compression parameters. In the initial integration stage of the model, only the calibration set is used to try different distance metrics to initially experience the quantization results; if the effect of the fast method does not meet the requirements, the optimal quantization parameter search and configuration are performed to obtain the best quantization accuracy, and on this basis, the parameters are adjusted to make the final performance and accuracy balance (a validation set is required); if it is still difficult to achieve a good balance between performance and effect, further fine-tune the maximum and minimum values ​​and mixed precision quantization. Since the performance and accuracy of the model after quantization is related to multiple factors, this is a gradual optimization process.

[0008] Method for format conversion based on smart energy unit model deployment and test: Convert the model format into a format that can run on the smart energy unit operating system. By considering compatibility and expansion requirements, better performance acceleration and more general cross-platform performance can be obtained in the actual deployment phase. For example, the trained model is converted into an intermediate representation file through a model optimizer. At the same time, the execution efficiency of the neural network topology structure is optimized during the model conversion process to improve model efficiency and performance.

[0009] Model packaging method: The model packaging method is reflected in the form of a library on the operating system for loading and calling by the operating system or application. The interface functions implemented include model loading (loading the format-converted model to the system), model running (starting and running the model, combining with the NPU unit to run the model operator through the NPU), model data input (realizing the input of historical net load data, weather data, date data and other factors into the model), and model output acquisition (obtaining the load forecast results of the model).

[0010] Test method: According to the characteristics of the test data and the prediction requirements, determine the appropriate evaluation indicators to evaluate the model prediction results. The evaluation results can be used as a reference to optimize and adjust the model, thereby providing support for various demand response resources on the user side to actively participate in the operation and regulation of the power market.

[0011] The beneficial effects of the present invention are: 1. Improving the compatibility of the model with the smart energy unit

[0012] Optimizing resource utilization Through in-depth learning of the hardware and software requirements for model deployment and the resource matching of the smart energy unit, the net load forecasting model can be accurately adapted to the limited resource environment of the smart energy unit. For example, the calculation support, memory, and storage space required by the model are reasonably evaluated to ensure that the model runs normally without wasting the hardware resources of the smart energy unit, thereby improving the utilization of hardware resources.

[0013] The adaptation method between the model and the smart energy unit enables the model to fully call upon the operating system and hardware resources of the smart energy unit, including the computing power of the NPU, thus avoiding performance bottlenecks caused by resource mismatch and improving the operating efficiency of the entire smart energy unit system.

[0014] Enhance system compatibility Perform necessary configuration and driver development on the operating system and ensure that the model dependency files match the operating system, which helps improve the compatibility of the net load forecasting model on different types of smart energy unit operating systems. Whether it is the existing operating system version or the possible future upgraded version, this method can make the model run stably and reduce the risk of model failure due to system updates.

[0015] 2. Improving Model Performance

[0016] The application of compression methods effectively reduces the size of the net load prediction model. The optimal quantization compression strategy is determined by analyzing the impact of different quantization strategies (such as dynamic quantization, mixed precision quantization, etc.) on model performance and storage, and combined with the NPU unit of the smart energy unit for compression processing, so that the model can be smoothly deployed and run on a resource-limited device such as the smart energy unit. This not only saves storage space, but also reduces the memory usage of the model during runtime, and improves the loading and running speed of the model.

[0017] The model optimization method minimizes the selected metrics layer by layer in the hyperspace to obtain the optimal quantitative compression parameters. By gradually tuning the model in stages, from the initial integration stage to the optimal quantitative parameter search and the final fine-tuning process, it can further improve the quantitative accuracy of the model while ensuring the model performance, and improve the prediction accuracy of the net load forecasting model, thereby providing more accurate load forecasting data for the operation and regulation of the power market.

[0018] Taking advantage of heterogeneous computing Based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit, the computing power advantage of the NPU for the model algorithm is combined with the compression optimization method. The efficient computing power of the NPU can accelerate the calculation process of the model, especially when dealing with complex computing tasks such as net load forecasting, which significantly improves the computing efficiency of the model, shortens the prediction time, and enables the net load forecast to provide support for the decision-making of the power market in a more timely manner.

[0019] 3. Achieve Effective Model Deployment and Testing

[0020] The convenient model deployment format conversion method converts the model into the operating format supported by the smart energy unit system, taking into account compatibility and expansion requirements, and obtaining better performance acceleration and more general cross-platform performance in the actual deployment stage. This makes it easy to deploy the net load forecasting model on the smart energy unit, and has good scalability, which is convenient for subsequent adjustments based on the upgrade or new function requirements of the smart energy unit.

[0021] The model encapsulation method designs a clear and easy-to-use logical interface, including functions such as model loading, running, and data input and output processing. This enables the operating system or application to easily call and use the converted model, reducing the difficulty of integrating the model into the smart energy system and improving development efficiency.

[0022] Reliable model evaluation and optimization The test method of the prediction model deployed on the smart energy unit selects appropriate evaluation indicators to evaluate the model prediction results according to the characteristics of the test data and the prediction requirements. This scientific and reasonable evaluation method can accurately reflect the performance of the model. The evaluation results are fed back to the model training process and can be used to adjust the model parameters and optimize the model's prediction ability. Through continuous evaluation and optimization, the net load forecasting model can better adapt to different application scenarios, improving the reliability and practicality of the model in power market operation and regulation.

[0023] 4. Help improve the power market operation and regulation capabilities

[0024] The method of the present invention provides strong support for various user-side demand response resources to actively participate in the operation and regulation of the power market. By efficiently deploying and optimizing the net load prediction model on the smart energy unit, the net load situation can be predicted more accurately, so that demand response resources (such as distributed energy resources, adjustable loads, etc.) can reasonably arrange production, consumption or energy storage activities according to the prediction results, thereby improving the flexibility and stability of the power market.

[0025] Improving the scientific nature of power market decision-making Accurate net load forecasting helps power market participants (such as power generation companies, grid operators, power sales companies, etc.) to formulate more scientific and reasonable operation strategies. For example, power generation companies can adjust power generation plans based on the predicted net load, grid operators can optimize grid dispatch, and power sales companies can formulate more accurate power sales price strategies, thereby improving the operating efficiency and economic benefits of the entire power market. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 : Process flow chart of the present invention. DETAILED DESCRIPTION

[0027] Deploying the prediction model on the edge device enables real-time data processing and analysis. Since the edge device is close to the data source, it can reduce the delay of data transmission, allowing the power system to respond to load changes more quickly and adjust the operation strategy in time. The traditional centralized data processing model may cause the server to be overloaded, especially when processing a large amount of real-time data. By deploying the load prediction model on the edge device, part of the data processing work can be completed locally, which effectively reduces the burden on the central server. Processing data on the edge device can reduce the dependence on the central server and reduce the risk of data being stolen or tampered with during transmission. In addition, even in the case of network instability or interruption, the edge device can still rely on the local model to make predictions to ensure the stable operation of the power system. Edge computing can reduce the demand and energy consumption of data transmission. Since the data is processed locally, there is no need to transmit all the data to the remote server, which reduces energy consumption and carbon emissions and is more environmentally friendly. With the development of smart grids, the amount of data that the power system needs to process will continue to grow. Deploying the load prediction model on the edge device can provide better scalability and flexibility, because the number of edge devices can be increased or decreased as needed to adapt to changes in the amount of data. In summary, the quantified load forecasting model deployed on edge devices has multiple meanings, including real-time response, reducing server burden, improving data security, energy saving and environmental protection, and providing scalability and flexibility. These advantages will help the power system operate more efficiently, safely, and environmentally friendly.

[0028] Therefore, the present invention deploys the net load prediction model on the smart energy unit and combines it with the high-performance main control NPU in the smart unit to further improve the performance of the model.

[0029] First, combined with the needs of the model and the resource situation of the smart energy unit, the compatibility and adaptation mechanism of the smart energy unit and the model, and secondly, the compression and optimization methods of the model, in order to adapt to the relatively limited resources of the smart energy unit, and explore the model optimization strategy for the energy unit main control chip and the NPU unit therein. Finally, based on the model deployment and testing of the smart energy unit, the model is converted into the operating format supported by the smart energy unit system, and the encapsulation method is used to integrate the model into the system's application program, so as to facilitate the active participation of various demand response resources on the user side in the operation and regulation of the power market. At the same time, combined with the load forecasting scenario, a suitable method is selected to test the model and evaluate the performance of the model in typical application scenarios.

[0030] Compatibility and adaptation mechanism between smart energy units and models

[0031] First, the software and hardware requirements for model deployment. By learning the complexity of the model, we evaluate the required example support, as well as the memory and storage space requirements, to ensure the process flow after the model is deployed. Determine whether additional dependency files are required based on the model and framework used. Analyze the security requirements for model deployment, such as isolation access control. In short, the software and hardware requirements for model deployment depend on multiple factors, including model complexity, data volume, performance requirements, and budget. Various factors need to be considered comprehensively to ensure the success and efficiency of deployment. Secondly, we learn the resource matching degree of smart energy units, the operating system and hardware of smart energy, especially the support of NPU for model examples, and whether it can meet the load forecasting requirements. Finally, the operating system is the basis for model deployment. For specific hardware requirements, corresponding driver support is required, and for special dependency files, they need to match the operating system. Therefore, an adaptation method for the prediction model and the smart energy unit is required to realize the model's call to software and hardware resources.

[0032] Model compression and optimization method of heterogeneous computing chip based on smart energy unit CPU+NPU

[0033] With the continuous development of artificial intelligence, neural network models have gradually become an important method in the field of power load forecasting. However, the complexity and resource requirements of these models have also increased. The operation of the prediction model requires a lot of computing resources and storage space. This greatly increases the cost of deploying and running the model, especially on resource-limited devices such as smart energy units. It becomes difficult or even impossible to deploy. Therefore, model compression and optimization methods have become key content. First, the compression of the learning model aims to reduce the size of the model so that it can be deployed and run in an environment with limited resources. Secondly, the learning model optimization method aims to improve the performance of the model in order to obtain better performance in a given resource environment. The prediction model of this project runs on a smart energy unit, and the heterogeneous computing chip on it contains an NPU unit. How to combine the computing power advantage of the NPU for the model algorithm with the compression optimization method can further improve the performance of the model.

[0034] Model deployment and testing based on smart energy units

[0035] First, how to convert the trained model from one format to a format that can run on the smart energy unit operating system, and convert the trained model into an intermediate representation file through the model optimizer, so as to obtain better performance acceleration and more general cross-platform performance in the actual deployment stage. Not only can it run on the smart energy unit operating system, but it also has certain cross-platform characteristics. During the model conversion process, the execution efficiency of the neural network topology structure is optimized to improve the efficiency and performance of the model. Secondly, the model encapsulation method, design clear and easy-to-use logical interfaces, so that applications or operating systems can easily call and use the converted models. The implemented interface functions need to include model loading, model running, model data input and output processing and other functions. Finally, the test method of the prediction model deployed on the smart energy unit, according to the characteristics of the test data and the prediction requirements, selects appropriate evaluation indicators to evaluate the model prediction results. The evaluation results can be fed back to the model training process to adjust the model parameters to optimize the model's prediction ability.

[0036] Compatibility of smart energy units and models

[0037] First, the model deployment software and hardware requirements are determined based on the model complexity, data volume, performance requirements, and budget. Second, the smart energy unit operating system and hardware resources are evaluated to determine whether they meet the requirements. Finally, the operating system is configured and driver developed as necessary, and dependent files are deployed on the model to meet the model requirements.

[0038] Model compression and optimization based on the characteristics of CPU+NPU heterogeneous computing chips of smart energy units

[0039] Model compression and optimization is a quantization process of a model. First, the compression method of the model affects the performance and storage of the model through different quantization strategies. Dynamic quantization, mixed precision quantization and other algorithms reduce the accuracy loss of the model, determine the quantization compression strategy, and compress the model based on this strategy combined with the NPU unit on the heterogeneous computing chip. Secondly, the model is optimized by minimizing the selected metric layer by layer in a hyperspace composed of distribution and feasible maximum and minimum distribution, and finally obtaining the optimal quantization compression parameters. The performance accuracy of the model after quantization is related to multiple factors, so it is a step-by-step optimization process. The model needs to be tuned step by step according to the progress of the project according to the following steps: In the initial integration stage of the model, only the calibration set is required, not the verification set, and different distance metrics are tried to experience the quantization results. If the effect of the fast method cannot meet the requirements, the optimal quantization parameter search and configuration are performed to obtain the best quantization accuracy. On this basis, the parameters are adjusted to make the final performance and accuracy balance. This method requires the preparation of a verification set. If it is still difficult to obtain a good balance between performance and effect, it is necessary to further implement fine-tuning of the maximum and minimum values ​​and mixed precision quantization.

[0040] Deployment and testing based on smart energy unit model

[0041] First, the format conversion method converts the format of the model into a format that can run on the smart energy unit operating system, and considers compatibility and expansion requirements to obtain better performance acceleration and more general cross-platform performance in the actual deployment stage. Secondly, the model encapsulation method is reflected in the form of a library on the operating system for loading and calling by the operating system or application. The implemented interfaces include model loading: loading the format-converted model onto the system; model running: starting and running the model, and combining it with the NPU unit to run the model operator through the NPU; model data input: inputting data of historical net load data, weather data, date data and other factors into the model; model output acquisition: obtaining the load forecast results of the model. According to the characteristics of the test data and the forecast requirements, determine the appropriate evaluation indicators to evaluate the model forecast results. The evaluation results can be used as a reference for optimizing the model. Through the deployment and testing of the model, it is possible to provide support for various demand response resources on the user side to actively participate in the operation and regulation of the power market.

[0042] Example 1, implementation of compatible adaptation of smart energy units and models First, analyze the structure and complexity of the net load prediction model in detail. For example, for a net load prediction model based on deep learning, the number of layers of the model, the number of neurons in each layer and other parameters are statistically analyzed, and the number of examples, memory usage and storage space required for the model to run are estimated based on these parameters. At the same time, based on the framework used by the model (such as TensorFlow or PyTorch), determine whether additional dependent files, such as specific library files or plug-ins, are required. For the security requirements of the model, set access permissions, such as using encryption methods to ensure the security of the model file, and set the access levels of different users or modules to the model.

[0043] Secondly, conduct a comprehensive evaluation of the smart energy unit. Obtain the operating system type (such as Linux-based system) and its version information of the smart energy unit, and check the hardware configuration, including CPU model, memory capacity, and NPU performance parameters (such as computing power, cache size, etc.). Through actual testing or theoretical calculations, evaluate the NPU's support capabilities for the net load prediction model example, for example, by running some simple model examples to observe the NPU's execution speed and resource usage.

[0044] Finally, perform the adaptation operation. According to the hardware situation of the smart energy unit, develop or install the corresponding driver, such as developing an adaptive driver for a specific model of NPU to ensure that the operating system can correctly identify and utilize the functions of the NPU. During the model deployment process, correctly install the required dependency files on the smart energy unit, and perform necessary configuration on the operating system, such as setting environment variables, so that the model can successfully call the software and hardware resources of the smart energy unit.

[0045] Model compression and optimization implementation based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit Model compression implementation: In the model compression stage, different quantization strategies are first tried for the net load prediction model. For example, for the dynamic quantization strategy, some sample data are selected to quantize the model, and the accuracy loss and storage space reduction of the model are observed. For mixed precision quantization, the accuracy parameters of different layers are adjusted to compare the performance of the model under different parameter settings. Through multiple experiments, the optimal quantization compression strategy is determined. Then, combined with the NPU unit of the smart energy unit, the model is further compressed and adjusted according to the computing characteristics of the NPU, such as formatting the parameters of the model according to the data format requirements of the NPU.

[0046] Model optimization implementation: During the model optimization process, first determine the indicators used to measure model performance, such as mean square error (MSE) or mean absolute error (MAE). In the initial integration stage of the model, only the calibration set data is used, for example, a part of the historical net load data is selected as the calibration set. Try different distance measurement methods, such as Euclidean distance or Manhattan distance, and observe the quantization results of the model. If the preliminary results do not meet the requirements, for example, the prediction error is large, search for the optimal quantization parameters. Use the validation set data (another part is selected from the historical data) to find the parameters with the best quantization accuracy through a search algorithm (such as a genetic algorithm or a particle swarm optimization algorithm). On this basis, adjust the model parameters to balance performance and accuracy. If satisfactory results are still not achieved, further fine-tune the maximum and minimum values ​​and mixed precision quantization parameters, such as adjusting the maximum and minimum ranges of quantization according to the distribution characteristics of the data, and optimizing the settings of mixed precision.

[0047] Implementation of format conversion based on smart energy unit model deployment and testing: For the net load prediction model, use the model optimizer to convert it from the original training format (such as the common .h5 format) to an intermediate representation file (such as the ONNX format). During the conversion process, analyze the compatibility requirements of the smart energy unit operating system and possible future expansion requirements. For example, if the operating system of the smart energy unit may be upgraded to a version that supports more instruction sets, this compatibility should be taken into account during format conversion. At the same time, optimize the execution efficiency of the neural network topology, such as by merging some computing layers or simplifying the network connection method to improve the running efficiency of the model on the smart energy unit.

[0048] Model encapsulation implementation: Design the model encapsulation library and build the library file on the operating system of the smart energy unit. Implement the model loading function and write code to correctly read the converted model file and load it into the memory; the model running function is implemented through the interactive code with the NPU unit, and the model operators are assigned to the NPU for execution to increase the calculation speed; the model data input function can receive historical net load data, weather data, date data and other formats of data and convert them into a format acceptable to the model by writing a data interface; the model output acquisition function can correctly extract the model's prediction results and perform necessary format conversion.

[0049] Test implementation: According to the characteristics of net load forecasting, select appropriate evaluation indicators, such as accuracy, recall, F1-score, etc. For test data, select historical net load data of different time periods and seasons, as well as corresponding weather data and date data. Input the test data into the model deployed on the smart energy unit, obtain the forecast results, and then evaluate according to the selected evaluation indicators. According to the evaluation results, if the forecast effect is not good, adjust the model parameters (such as the weight and bias of the neural network) or optimize the model structure (such as increasing or decreasing the number of layers) to improve the model's forecasting ability.

Claims

1. A method for deploying and optimizing a net load prediction model on a smart energy unit, characterized in that: include: S1. Compatibility and adaptation of smart energy units and models; Specifically, it includes: evaluating the required instance support, memory and storage space requirements by analyzing the model complexity, data volume, performance requirements and budget model deployment software and hardware requirements, determining whether additional dependency files are needed and learning security requirements based on the model and framework; Evaluate the operating system and hardware resources of the smart energy unit, focusing on whether the NPU's support for model examples meets the load forecasting requirements; The adaptation method of the prediction model and the smart energy unit, evaluates the operating system and hardware resources of the smart energy unit to determine whether they meet the requirements, performs necessary configuration and driver development on the operating system, and deploys dependent files on the model to meet the model requirements; Model compression and optimization based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit, including: model compression method, by exploring the impact of different quantization strategies on model performance and storage, determining the quantization compression strategy and combining the NPU unit to compress the model; S2, model optimization method, which minimizes the selected metric layer by layer in the hyperspace composed of distribution and feasible maximum and minimum distribution to obtain the optimal quantitative compression parameters, and gradually tunes the model in stages according to the progress of the project; Based on the deployment and testing of the smart energy unit model, it specifically includes: format conversion method, which converts the model format into a format that can be run on the smart energy unit operating system, and obtains better performance acceleration ratio and cross-platform performance in the actual deployment stage by considering compatibility and expansion requirements; S3, the model packaging method, which is reflected in the form of a library on the operating system to realize interface functions such as model loading, running, data input and output acquisition; S4. Testing method: Determine appropriate evaluation indicators to evaluate the model prediction results based on the characteristics of the test data and prediction requirements. The evaluation results can be used to optimize and adjust the model.

2. The method for deploying and optimizing the net load prediction model on a smart energy unit according to claim 1, characterized in that: In the compatible adaptation of the smart energy unit and the model, the security requirements for the model deployment include isolation access control.

3. The method for deploying and optimizing the net load prediction model on a smart energy unit according to claim 1, characterized in that: In the model compression and optimization based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit, the different quantization strategies include dynamic quantization and mixed precision quantization algorithms.

4. The method for deploying and optimizing the net load prediction model on a smart energy unit according to claim 1, characterized in that: In the deployment and testing of the smart energy unit model, the format conversion method includes converting the trained model into an intermediate representation file through a model optimizer, and optimizing the execution efficiency of the neural network topology structure during the conversion process.

5. The method for deploying and optimizing the net load prediction model on a smart energy unit according to claim 1, characterized in that: In the deployment and testing of the smart energy unit model, in the interface function implemented by the model encapsulation method, the model running function includes combining with the NPU unit to run the model operator through the NPU.

6. The method for deploying and optimizing the net load prediction model on a smart energy unit according to claim 1, characterized in that: In the deployment and testing of the smart energy unit model, the evaluation indicators selected in the testing method include one or more of accuracy, recall rate, and F1-score.

Images