A method for deploying and optimizing a net load prediction model on a smart energy unit

Abstract

This invention relates to a method for deploying and optimizing a net load forecasting model on a smart energy unit. The method includes: ensuring compatibility and adaptation between the smart energy unit and the model, specifically including: studying the hardware and software requirements for model deployment based on model complexity, data volume, performance requirements, and budget; assessing the required computational support, memory, and storage space requirements by analyzing model complexity; determining whether additional dependency files are needed based on the model and framework and studying security requirements; employing an ensemble learning algorithm for target detection; and introducing a time-series event model with a self-attention mechanism for fault warning and troubleshooting guidance, predicting the probability of potential faults in key equipment. The beneficial effects of this invention are: 1. Improving the compatibility between the model and the smart energy unit. 2. Enhancing model performance. 3. Achieving effective model deployment and testing. 4. Contributing to improved power market operation and regulation capabilities.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, specifically relating to a method for deploying and optimizing a net load forecasting model on a smart energy unit. Background Technology

[0002] With the development of smart energy systems, net load forecasting plays a crucial role in the operation and regulation of the electricity market. Neural network models are widely used in net load forecasting, but these models are highly complex and resource-intensive. Smart energy units have relatively limited resources, and deploying net load forecasting models on them faces numerous challenges, such as model compatibility with smart energy units and operational efficiency under limited resources. Currently, there is a lack of systematic methods to effectively deploy and optimize the performance 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 electricity market. Summary of the Invention

[0003] To address the technical problems in the background, this invention proposes a method for deploying and optimizing a net load forecasting model on a smart energy unit. The method considers the hardware and software requirements for model deployment, taking into account factors such as model complexity, data volume, performance requirements, and budget. It analyzes model complexity to assess the required computational support, memory, and storage space requirements, ensuring the normal operation of the process after model deployment. Based on the model and framework used, it determines whether additional dependency files are needed and learns the security requirements for model deployment, such as isolation access control.

[0004] Smart Energy Unit Resource Matching Assessment: In-depth study of the smart energy unit's operating system and hardware, with a focus on whether the NPU's support for model calculations can meet the load forecasting requirements.

[0005] Adaptation Method: Since the operating system is the foundation for model deployment, specific hardware requirements necessitate corresponding drivers, and special dependency files must match the operating system. Therefore, an adaptation method for the predictive model and smart energy unit is needed to enable the model to utilize hardware and software resources. Specifically, the operating system and hardware resources of the smart energy unit are evaluated to determine if they meet the requirements. Then, the operating system is configured as necessary, and drivers are developed. Finally, dependency files are deployed on the model to meet its requirements.

[0006] Model compression and optimization based on the characteristics of CPU+NPU heterogeneous computing chips in smart energy units: The model compression method explores the impact of different quantization strategies (such as dynamic quantization, mixed precision quantization, etc.) on model performance and storage, reduces the accuracy loss of the model, determines the quantization compression strategy, and combines the NPU unit on the heterogeneous computing chip to compress the model.

[0007] Model optimization method: The selected metric is minimized layer by layer in a hyperspace composed of the distribution and feasible maximum-minimum distributions to obtain the optimal quantization compression parameters. In the initial model integration phase, only the calibration set is used to try different distance metrics to initially experience the quantization results. If the fast method does not meet the requirements, the optimal quantization parameters are searched and configured to obtain the best quantization accuracy. Based on this, the parameters are adjusted to achieve a final balance between performance and accuracy (requiring a validation set). If a good balance between performance and accuracy is still difficult to achieve, further fine-tuning of maximum-minimum and mixed-precision quantization is implemented. Since the post-quantization performance and accuracy of the model are related to multiple factors, this is a gradual optimization process.

[0008] The method based on the deployment and testing format conversion of smart energy unit models involves converting the model's format into one that can run on the smart energy unit operating system. By considering compatibility and scalability requirements, a better performance speedup and more universal cross-platform performance can be achieved in the actual deployment phase. For example, the trained model can be converted into an intermediate representation file through a model optimizer. At the same time, the execution efficiency of the neural network topology can be optimized during the model conversion process to improve model efficiency and performance.

[0009] Model encapsulation method: This refers to the method of encapsulating the model and implementing it as a library on the operating system for loading and calling by the operating system or applications. Its implemented interface functions include model loading (loading the converted model onto the system), model running (starting and running the model, and combining with the NPU unit to run model operators), model data input (inputting historical net load data, weather data, date data, and other factors into the model), and model output acquisition (obtaining the model's load forecast results).

[0010] Testing Method: Based on the characteristics of the test data and the forecast requirements, appropriate evaluation indicators are determined to evaluate the model's forecast results. The evaluation results can be used as a reference to optimize and adjust the model, thereby providing support for users to actively participate in the operation and regulation of the power market.

[0011] The beneficial effects of this invention are: 1. Improved compatibility between the model and smart energy units.

[0012] Optimizing resource utilization involves deeply learning the hardware and software requirements of the model deployment and the resource matching degree of the smart energy unit. This allows for the accurate adaptation of the net load forecasting model to the limited resource environment of the smart energy unit. For example, by reasonably assessing the computational support, memory, and storage space required by the model, it ensures that the model operates normally without wasting the hardware resources of the smart energy unit, thereby improving the utilization rate of hardware resources.

[0013] The model-smart energy unit adaptation method enables the model to fully utilize the smart energy unit's operating system and hardware resources, including the NPU's computing power, avoiding performance bottlenecks caused by resource mismatch and improving the overall operating efficiency of the smart energy unit system.

[0014] Enhancing system compatibility involves necessary configuration and driver development for the operating system, ensuring that model dependency files match the operating system. This helps improve the compatibility of the net load forecasting model across different types of smart energy unit operating systems. Regardless of the existing operating system version or potential future upgrades, this method ensures stable model operation and reduces the risk of model failure due to system updates.

[0015] II. Improving Model Performance

[0016] The application of model compression and optimization methods effectively reduces the size of the net load prediction model. By analyzing the impact of different quantization strategies (such as dynamic quantization and mixed precision quantization) on model performance and storage, the optimal quantization compression strategy is determined and combined with the NPU unit of the smart energy unit for compression processing. This enables the model to be successfully deployed and run on resource-constrained devices like smart energy units. This not only saves storage space but also reduces the memory footprint during model runtime, improving the model's loading and running speed.

[0017] The model optimization method minimizes selected metrics layer by layer in the hyperspace to obtain the optimal quantization compression parameters. By gradually optimizing the model in stages, from the initial integration stage to the search for the optimal quantization parameters and the final fine-tuning process, the quantization accuracy of the model can be further improved while ensuring model performance. This enhances 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] Leveraging the advantages of heterogeneous computing, this approach combines the CPU+NPU heterogeneous computing chip of the smart energy unit with compression optimization methods, utilizing the NPU's computational power for model algorithms. The NPU's high-efficiency computing capabilities accelerate the model's computation process, especially when handling complex computational tasks such as net load forecasting. This significantly improves the model's computational efficiency, shortens forecasting time, and enables net load forecasting to provide more timely support for electricity market decision-making.

[0019] III. Achieving Effective Model Deployment and Testing

[0020] A convenient model deployment format conversion method transforms the model into an operating format supported by the smart energy unit system. Considering compatibility and scalability requirements, it achieves superior performance speedup and more versatile cross-platform performance during actual deployment. This enables the net load forecasting model to be easily deployed on smart energy units with good scalability, facilitating subsequent adjustments based on upgrades or new feature requirements of the smart energy unit.

[0021] The model encapsulation method features a clear and easy-to-use logical interface, including functions for model loading, running, and data input / output processing. This allows operating systems or applications to easily call and use the transformed model, reducing the difficulty of integrating the model into the smart energy system and improving development efficiency.

[0022] A reliable model evaluation and optimization method for testing predictive models deployed on smart energy units involves selecting appropriate evaluation indicators based on the characteristics of the test data and predictive requirements to assess the model's prediction results. This scientific and reasonable evaluation method accurately reflects the model's performance, and the evaluation results are fed back into the model training process, allowing for the adjustment of model parameters and optimization of the model's predictive capabilities. Through continuous evaluation and optimization, the net load prediction model can better adapt to different application scenarios, improving its reliability and practicality in power market operation and regulation.

[0023] IV. Helping to improve the operation and regulation capabilities of the electricity market

[0024] This invention provides strong support for various demand response resources on the user side to actively participate in the operation and regulation of the electricity market. By efficiently deploying and optimizing net load forecasting models on smart energy units, net load conditions can be predicted more accurately. This allows demand response resources (such as distributed energy resources and adjustable loads) to rationally arrange production, consumption, or energy storage activities based on the forecast results, thereby improving the flexibility and stability of the electricity market.

[0025] Improving the scientific rigor of electricity market decision-making and ensuring accurate net load forecasting helps electricity market participants (such as power generation companies, grid operators, and electricity retailers) formulate more scientific and rational operating strategies. For example, power generation companies can adjust their generation plans based on the forecasted net load, grid operators can optimize grid dispatching, and electricity retailers can develop more precise electricity pricing strategies, thereby improving the overall operational efficiency and economic benefits of the electricity market. Attached Figure Description

[0026] Figure 1 : Process flow diagram of this invention. Detailed Implementation

[0027] Deploying predictive models on edge devices enables real-time data processing and analysis. Because edge devices are located close to the data source, data transmission latency is reduced, allowing the power system to respond more quickly to load changes and adjust operating strategies more promptly. Traditional centralized data processing models can overload servers, especially when processing large amounts of real-time data. By deploying load forecasting models on edge devices, some data processing can be done locally, effectively reducing the burden on central servers. Processing data on edge devices reduces reliance on central servers and lowers the risk of data theft or tampering during transmission. Furthermore, even in the event of network instability or outages, edge devices can still rely on local models for forecasting, ensuring the stable operation of the power system. Edge computing reduces data transmission requirements and energy consumption. Since data is processed locally, it is not necessary to transmit all data to remote servers, thus reducing energy consumption and carbon emissions, making it 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 load forecasting models on edge devices provides better scalability and flexibility, as the number of edge devices can be increased or decreased as needed to adapt to changes in data volume. In summary, deploying quantified load forecasting models on edge devices offers multiple benefits, including real-time response, reduced server load, improved data security, energy conservation and environmental protection, and scalability and flexibility. These advantages will contribute to more efficient, safe, and environmentally friendly operation of the power system.

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

[0029] First, considering the model's requirements and the resource availability of the smart energy unit, the compatibility and adaptation mechanisms between the smart energy unit and the model are determined. Second, model compression and optimization methods are explored to adapt to the relatively limited resources of the smart energy unit, and optimization strategies for the energy unit's main control chip and its NPU unit are investigated. Finally, based on the deployment and testing of the model on the smart energy unit, the model is converted into an operating format supported by the smart energy unit system, and encapsulation methods are developed to integrate the model into the system's applications. This facilitates support for users to actively participate in the operation and regulation of the electricity market by responding to various demand-side factors. Simultaneously, in conjunction with load forecasting scenarios, appropriate methods are selected to test the model and evaluate its performance in typical application scenarios.

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

[0031] First, the hardware and software requirements for model deployment are assessed by learning from the model's complexity, evaluating the necessary computational support, and determining memory and storage requirements to ensure smooth workflow after deployment. Based on the model and framework used, it's determined whether additional dependency files are needed. Security requirements for model deployment, such as isolation access control, are also analyzed. In short, the hardware and software requirements for model deployment depend on multiple factors, including model complexity, data volume, performance requirements, and budget. A comprehensive consideration of these factors is necessary to ensure successful and efficient deployment. Second, the resource compatibility of the smart energy unit is learned, including the smart energy operating system and hardware, especially the NPU's support for model computations, to ensure it meets load forecasting requirements. Finally, the operating system is the foundation of model deployment. Specific hardware requirements necessitate corresponding driver support, and special dependency files need to be compatible with the operating system. Therefore, a method for adapting the forecasting model and smart energy unit is needed to enable the model to utilize hardware and software resources.

[0032] Compression and Optimization Methods Based on Smart Energy Unit CPU+NPU Heterogeneous Computing Chip Model

[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, and the operation of forecasting models requires a large amount of computing resources and storage space. This significantly increases the cost of deploying and running models, making deployment difficult or even impossible, especially on resource-constrained devices such as smart energy units. Therefore, model compression and optimization methods have become crucial. First, learning model compression aims to reduce the model size for deployment and operation in resource-limited environments. Second, learning model optimization methods aim to improve model performance to achieve better performance in a given resource environment. The forecasting model in this project runs on a smart energy unit, whose heterogeneous computing chip includes an NPU unit. Combining the computational advantages of the NPU for model algorithms with compression and optimization methods can further improve the model's performance.

[0034] Model Deployment and Testing Based on Smart Energy Units

[0035] First, how to convert the trained model from one format to one that can run on the smart energy unit operating system? This involves using a model optimizer to convert the trained model into an intermediate representation file, achieving better performance speedup and more versatile cross-platform performance during deployment. The model should not only run on the smart energy unit operating system but also possess certain cross-platform characteristics. This is achieved by optimizing the execution efficiency of the neural network topology during the model conversion process to improve model efficiency and performance. Second, the model encapsulation method involves designing clear and easy-to-use logical interfaces so that applications or operating systems can easily call and use the converted model. The implemented interface functions need to include model loading, model execution, and input / output processing of model data. Finally, the testing method for the prediction model deployed on the smart energy unit involves selecting appropriate evaluation metrics to evaluate the model's prediction results based on the characteristics of the test data and prediction requirements. The evaluation results can be fed back into the model training process to adjust model parameters and optimize the model's predictive capabilities.

[0036] Compatibility and Adaptation of Smart Energy Units and Models

[0037] First, the hardware and software requirements for model deployment are determined based on factors such as model complexity, data volume, performance requirements, and budget. Second, the operating system and hardware resources of the smart energy unit are evaluated to determine if they meet the requirements. Finally, the necessary configuration and driver development for the operating system are performed, and dependency files are deployed on the model to satisfy its requirements.

[0038] Model Compression and Optimization Based on the Characteristics of Heterogeneous Computing Chips (CPU+NPU) in Smart Energy Units

[0039] Model compression and optimization is a quantization process of a model. First, the compression method is considered, examining the impact of different quantization strategies on model performance and storage. Algorithms such as dynamic quantization and mixed-precision quantization reduce the accuracy loss of the model. A quantization compression strategy is determined, and based on this strategy, combined with NPU units on heterogeneous computing chips, the model is compressed. Second, model optimization is performed by progressively minimizing a selected metric within a hyperspace composed of the distribution and feasible maximum-minimum distributions, ultimately obtaining the optimal quantization compression parameters. The post-quantization performance and accuracy of the model are related to multiple factors, thus it is a gradual optimization process. The model needs to be gradually tuned according to the project progress, following these steps: In the initial model integration phase, only a calibration set is needed, without relying on a validation set. Different distance metrics are tried to initially assess the quantization results. If the fast method's performance does not meet the requirements, the optimal quantization parameters are searched and configured to obtain the best quantization accuracy. Based on this, the parameters are adjusted to achieve a final balance between performance and accuracy. This type of method requires a validation set. If a good balance between performance and accuracy is still difficult to achieve, further fine-tuning of maximum-minimum distributions and mixed-precision quantization is required.

[0040] Deployment and Testing Based on Smart Energy Unit Model

[0041] First, a format conversion method transforms the model's format into one that can run on the smart energy unit's operating system. This considers compatibility and scalability requirements to achieve superior performance speedup and more versatile cross-platform performance during actual deployment. Second, a model encapsulation method is implemented, presenting the model as a library on the operating system for loading and invocation by the operating system or applications. The implemented interfaces include: model loading: loading the converted model onto the system; model execution: starting and running the model, integrating with the NPU unit to run model operators; model data input: inputting data from various factors such as historical net load data, weather data, and date data into the model; and model output acquisition: obtaining the model's load forecast results. Based on the characteristics of the test data and forecasting requirements, appropriate evaluation indicators are determined to evaluate the model's forecast results. The evaluation results can be used as a reference for model optimization and adjustment. Through model deployment and testing, support is provided for various demand response resources on the user side to actively participate in the operation and regulation of the electricity market.

[0042] Example 1: Compatibility and Adaptation Implementation of Smart Energy Units and Models. First, a detailed analysis of the structure and complexity of the net load forecasting model is conducted. For example, for a deep learning-based net load forecasting model, parameters such as the number of layers and the number of neurons per layer are statistically analyzed. Based on these parameters, the number of computational instances, memory usage, and storage space required for the model's runtime are estimated. Simultaneously, based on the framework used by the model (such as TensorFlow or PyTorch), it is determined whether additional dependency files, such as specific library files or plugins, are needed. Regarding the model's security requirements, access permissions are set, such as using encryption methods to ensure the security of model files and setting different access levels for different users or modules.

[0043] Secondly, a comprehensive evaluation of the smart energy unit should be conducted. This includes obtaining the operating system type (e.g., Linux-based system) and version information of the smart energy unit, and checking the hardware configuration, including CPU model, memory capacity, and NPU performance parameters (e.g., computing power, cache size). Through actual testing or theoretical calculations, the NPU's ability to support net load forecasting model examples should be evaluated. For example, by running some simple model examples, the execution speed and resource consumption of the NPU can be observed.

[0044] Finally, adaptation is performed. Based on the hardware specifications of the smart energy unit, corresponding drivers are developed or installed, such as developing an adapted driver for a specific NPU model, ensuring the operating system can correctly recognize and utilize the NPU's functions. During model deployment, the necessary dependency files are correctly installed on the smart energy unit, and the operating system is configured as needed, such as setting environment variables, so that the model can smoothly access the smart energy unit's hardware and software resources.

[0045] Model Compression and Optimization Based on the Characteristics of Heterogeneous Computing Chips (CPU+NPU) in Smart Energy Units: Model Compression Implementation: In the model compression stage, for the net load prediction model, different quantization strategies were first tried. For example, for dynamic quantization, a subset of sample data was selected to quantize the model, and the loss of accuracy and reduction in storage space were observed. For mixed-precision quantization, the accuracy parameters of different layers were adjusted, and the model performance under different parameter settings was compared. Through multiple experiments, the optimal quantization compression strategy was determined. Then, combined with the NPU unit of the smart energy unit, the model was further compressed and adjusted according to the computing characteristics of the NPU, such as formatting the model parameters according to the data format requirements of the NPU.

[0046] Model optimization implementation: During model optimization, the first step is to determine the metrics used to measure model performance, such as mean squared error (MSE) or mean absolute error (MAE). In the initial model ensemble phase, only calibration set data is used, for example, a subset of historical net load data is selected as the calibration set. Different distance metrics, such as Euclidean distance or Manhattan distance, are tested, and the quantization results are observed. If the initial results do not meet the requirements, such as large prediction errors, an optimal quantization parameter search is performed. Using validation set data (an additional subset selected from historical data), a search algorithm (such as a genetic algorithm or particle swarm optimization algorithm) is used to find parameters with the best quantization accuracy. Based on this, model parameters are adjusted to balance performance and accuracy. If satisfactory results are still not achieved, the maximum and minimum values ​​and mixed precision quantization parameters are further fine-tuned, for example, adjusting the maximum and minimum quantization ranges according to the data distribution characteristics, while simultaneously optimizing the mixed precision settings.

[0047] The deployment and testing of the smart energy unit model involved format conversion: For the net load prediction model, a model optimizer was used 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, the compatibility requirements and potential future expansion needs of the smart energy unit's operating system were analyzed. For example, if the smart energy unit's operating system might be upgraded to a version supporting more instruction sets, this compatibility must be considered during the format conversion. Simultaneously, the execution efficiency of the neural network topology was optimized, for example, by merging some computational layers or simplifying network connections, to improve the model's running efficiency on the smart energy unit.

[0048] Model encapsulation and implementation: Design a model encapsulation library and build the library files on the smart energy unit's operating system. Implement the model loading function, writing code to correctly read the converted model file and load it into memory; implement the model execution function through interaction code with the NPU unit, allocating the model's operators to the NPU for execution to improve computation speed; implement the model's data input function through a data interface, capable of receiving data in various formats such as historical net load data, weather data, and date data, and converting it into a format acceptable to the model; implement the model's output acquisition function to correctly extract the model's prediction results and perform necessary format conversions.

[0049] Test Implementation: Based on the characteristics of net load forecasting, select appropriate evaluation metrics, such as accuracy, recall, and F1-score. For test data, select historical net load data from different time periods and seasons, along with corresponding weather and date data. Input the test data into the model deployed on the smart energy unit, obtain the prediction results, and then evaluate them according to the selected evaluation metrics. Based on the evaluation results, if the prediction performance is unsatisfactory, adjust the model parameters (such as the weights and biases of the neural network) or optimize the model structure (such as adding or removing layers) to improve the model's predictive ability.

Claims

1. A method for deploying and optimizing a net load forecasting model on a smart energy unit, characterized in that, include: S1. Compatibility and adaptation of smart energy units and models; Specifically, this includes: assessing the required computational support, memory, and storage space requirements by analyzing the model complexity, based on the model's complexity, data volume, performance requirements, and budget for the model's deployment hardware and software requirements; determining whether additional dependency files are needed based on the model and framework; and learning security requirements. Assess the operating system and hardware resources of the smart energy unit, with a focus on evaluating whether the NPU's support for model calculations meets the load forecasting requirements. The method for adapting the prediction model and the smart energy unit involves evaluating the smart energy unit's operating system and hardware resources to determine if they meet the requirements, configuring the operating system and developing drivers as necessary, and deploying dependency files on the model to meet the model's requirements. Model compression and optimization based on the characteristics of the CPU+NPU heterogeneous computing chip of the smart energy unit, specifically including: model compression methods, by exploring the impact of different quantization strategies on model performance and storage, determining the quantization compression strategy and combining it with the NPU unit to compress the model; S2, Model optimization method, which minimizes the selected metric layer by layer in the hyperspace composed of the distribution and feasible maximum and minimum value distribution to obtain the optimal quantization compression parameter, and gradually optimizes the model in stages according to the progress of the project; Based on the deployment and testing of the smart energy unit model, the specific methods include: format conversion methods, which convert the model format into a format that can run on the smart energy unit operating system, and by taking into account compatibility and scalability requirements, a better performance speedup and cross-platform performance can be obtained in the actual deployment stage. S3, the model encapsulation method, is implemented in the form of a library on the operating system, realizing the interface functions of model loading, running, data input and output acquisition; S4. Testing Methods: Based on the characteristics of the test data and the prediction requirements, determine appropriate evaluation indicators to evaluate the model's prediction results. The evaluation results can be used to optimize and adjust the model.

2. The method for deploying and optimizing the net load forecasting model on a smart energy unit according to claim 1, characterized in that, In the compatibility and adaptation of smart energy units and models, the security requirements for the deployment of the models include isolated access control.

3. The method for deploying and optimizing the net load forecasting model on a smart energy unit according to claim 1, characterized in that, In 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 forecasting 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 during the conversion process.

5. The method for deploying and optimizing the net load forecasting 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 interface functions implemented by the model encapsulation method include the model running function, which combines with the NPU unit to run model operators through the NPU.

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

Images