Thermal generator set load prediction system based on machine learning

Abstract

The invention discloses a thermal generator set load prediction system based on machine learning, and the system overcomes the defects of a conventional load prediction method through an end-to-end solution composed of a comprehensive data collection module, a multi-level feature engineering module, a multi-model integration module, a self-adaptive optimization module and the like. The system collects multi-dimensional data from a thermal generator set control system, an environment monitoring system, a fuel detection system and a historical database, wherein the multi-dimensional data comprises load data, environment data, fuel characteristics and equipment operation data. According to the multi-dimensional data integration mode, the model not only depends on single historical load data, but combines multiple influence factors such as environment, fuel and equipment states, the problem that a traditional method depends on a single data source is solved, the load prediction precision is effectively improved, and the load prediction efficiency is improved. And the adaptive capacity of the prediction model to different environment conditions and equipment states is stronger.

Description

Technical Field

[0001] The present application relates to the technical field of load prediction of thermal power generation units based on machine learning, and in particular to a load prediction system of thermal power generation units based on machine learning. Background Art

[0002] The load forecasting of thermal power generating units is of great significance for the stable operation of power systems and the rational allocation of energy resources. Traditional load forecasting methods are mostly based on historical load data and simple linear regression or time series models, ignoring the impact of environmental factors, fuel characteristics and unit operating parameters on load. The forecasting effect is poor and it is difficult to adapt to the complex volatility of thermal power generation load.

[0003] In recent years, the application of machine learning and deep learning technologies in power load forecasting has gradually increased. However, the existing methods still have the following shortcomings:

[0004] There is a lack of comprehensive utilization of multi-dimensional data, such as the dynamic impact of environmental parameters and fuel characteristics on thermal power generation.

[0005] The traditional single model has low prediction accuracy and poor adaptability and cannot respond to load changes in real time.

[0006] The parameters of the prediction model are mostly fixed, making it difficult to automatically optimize them according to actual conditions, and the real-time adaptability of the model is limited. Summary of the invention

[0007] The present application provides a load forecasting system for thermal power generation units based on machine learning, aiming to solve the problem that the prior art lacks comprehensive utilization of multi-dimensional data.

[0008] A load forecasting system for thermal power generating units based on machine learning, the system includes data acquisition, data preprocessing, feature engineering, model selection and training, model optimization and prediction, result output and feedback modules, wherein:

[0009] Data acquisition module: obtains load data, environmental data, fuel characteristics and equipment operation data from the thermal power generator control system, environmental monitoring system, fuel detection system and historical database;

[0010] Data preprocessing module: detect and clean outliers, fill missing values ​​and standardize the collected data;

[0011] Feature engineering module: construct multi-level lag features, dynamic weighted environmental features, autoencoder dimensionality reduction features, cross features and periodic features based on the characteristics of thermal power generation load;

[0012] Model selection and training module: Multi-model integration based on deep learning and machine learning, including LSTM, CNN-LSTM hybrid model, XGBoost, random forest and time series baseline model, and hyperparameter optimization through Bayesian optimization and reinforcement learning;

[0013] Model optimization and prediction module: Use dynamic weighting strategy to integrate multiple models, optimize the long-term performance of the model based on adaptive feedback mechanism, and perform real-time load prediction;

[0014] Result output and feedback module: outputs the load forecast results to the dispatching system, and uses the actual load data for model performance monitoring and dynamic adjustment through the feedback mechanism.

[0015] In the above scheme, optionally, the environmental data collected by the data acquisition module include air temperature, humidity, and wind speed, and the weight of each environmental feature is dynamically adjusted in combination with the actual load demand of the thermal power generating unit.

[0016] In the above scheme, optionally, the multi-level lag features in the feature engineering module include intra-day lag features and inter-day lag features, which are used to capture short-term and long-term load change trends.

[0017] In the above solution, optionally, the feature engineering module uses an autoencoder to perform dimensionality reduction processing on multi-dimensional features to reduce the impact of noise and retain the core structural information of the data;

[0018] The model selection and training module includes a network structure based on a mixture of convolutional neural networks and long short-term memory networks. The CNN layer extracts local features in the input sequence, and the LSTM layer is used to capture long-term dependencies.

[0019] In the above scheme, optionally, the model selection and training module adopts a dynamically weighted multi-model integration strategy to dynamically adjust the weight of the model according to real-time error feedback to improve the overall prediction accuracy of the model.

[0020] In the above scheme, optionally, the model optimization and prediction module optimizes the hyperparameter settings of the model through a reinforcement learning algorithm, wherein the reinforcement learning agent automatically adjusts the hyperparameters of the model according to the prediction error feedback.

[0021] In the above scheme, optionally, the model optimization and prediction module includes an adversarial training mechanism based on GAN, the generator is used to output the load prediction value, and the discriminator feeds back an error signal to correct the prediction result of the generator, thereby reducing the error of load prediction.

[0022] In the above scheme, optionally, the model optimization and prediction module introduces a self-attention mechanism in the LSTM model, so that the model can dynamically allocate the importance of features at each time step, thereby improving the accuracy of load forecasting.

[0023] In the above scheme, optionally, the model optimization and prediction module performs load prediction based on the Transformer structure, which is used to capture long-term periodic relationships in load data and is suitable for predicting long-term load change trends.

[0024] In the above scheme, optionally, the result output and feedback module stores the predicted value in a time series database, and dynamically updates the prediction model in combination with the actual load data to maintain the adaptability of the model to different operating environments.

[0025] Compared with the prior art, this application has at least the following beneficial effects:

[0026] Based on further analysis and research on the problems of the prior art, this application recognizes that the prior art lacks the problem of comprehensive utilization of multi-dimensional data. Through an end-to-end solution consisting of modules such as comprehensive data collection, multi-level feature engineering, multi-model integration and adaptive optimization, it overcomes the shortcomings of traditional load forecasting methods and effectively solves the main problems in the background technology. This system collects multi-dimensional data from the control system, environmental monitoring system, fuel detection system and historical database of the thermal power generator set, including load data, environmental data, fuel characteristics and equipment operation data. This multi-dimensional data integration method makes the model not only rely on a single historical load data, but also combines multiple influencing factors such as environment, fuel, and equipment status, thereby overcoming the traditional method's dependence on a single data source. Through the fusion of multi-dimensional data, the model can more comprehensively and accurately characterize the changing trend of the generator set load, effectively improve the accuracy of load forecasting, and make the prediction model more adaptable to different environmental conditions and equipment status.

[0027] This system constructs multi-level lag features, dynamically weighted environmental features, autoencoder dimensionality reduction features, cross features and periodic features through feature engineering modules. Traditional load forecasting usually only considers the historical value of the load, while this method greatly improves the input information quality of the model by extracting multiple features (such as lag features and periodic features) and generating cross features. In particular, the dynamic weighting of environmental features and the dimensionality reduction of the autoencoder can effectively avoid the interference of redundant features on the model while ensuring the integrity of key features. By enhancing the diversity and importance of input features, the model can more accurately identify short-term fluctuations and long-term trends of the load, making the prediction more robust and more adaptable, thus solving the problem of low prediction accuracy caused by the lack of feature engineering support in traditional methods.

[0028] Traditional load forecasting methods mostly rely on a single model, such as a simple linear regression or time series model, while this solution adopts a multi-model integration strategy. The integrated model includes LSTM, CNN-LSTM hybrid model, XGBoost, random forest and other algorithms. Through the complementary advantages of different algorithms, the overall prediction effect of the system is improved. At the same time, through the dynamic weighted fusion method, the model weight is dynamically adjusted according to the actual error to adapt to different load modes. Through multi-model integration and dynamic weighting, the system can effectively overcome the limitations of a single model, maintain high accuracy under different load modes, significantly improve the stability and scope of application of load forecasting, and solve the problem of low prediction accuracy of a single model in the background technology.

[0029] To enhance the real-time optimization capability of the model, this solution introduces reinforcement learning and Bayesian optimization in the model optimization and prediction module to achieve automatic tuning of hyperparameters and dynamically adjust the prediction error through real-time feedback. The reinforcement learning agent automatically adjusts the model parameters according to the actual prediction effect, and Bayesian optimization further improves the adaptability of the model. The combination of reinforcement learning and Bayesian optimization enables the model to be adaptively optimized according to real-time feedback, improves the accuracy and adaptability of the model under different working conditions, solves the problem that traditional models cannot be dynamically adjusted according to actual needs, and makes the prediction system more intelligent and flexible.

[0030] The result output and feedback module stores the prediction results through the time series database, and introduces a feedback mechanism to update the model according to the actual load conditions to ensure that the model can always adapt to the real-time changing load environment. The system compares the real-time prediction results with the actual data, and dynamically adjusts the model according to the error size to maintain the stability and accuracy of the prediction effect. Through real-time feedback and model updates, the system can continuously optimize the prediction effect, effectively solving the limitations of the traditional load forecasting system that lacks feedback and self-optimization. This enables the model to always reflect the load demand of the generator set in actual applications, improving the long-term adaptability of the system.

[0031] Through the coordinated application of the above technical modules, this system not only improves the problems of single data source, insufficient feature extraction, poor model adaptability, etc. in traditional load forecasting, but also significantly improves the prediction accuracy, adaptive ability and long-term stability of the model, thus providing an efficient, intelligent and stable technical solution for thermal power generation load forecasting. This systematic and innovative load forecasting method effectively overcomes the shortcomings of the background technology and has significant application value and innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A block diagram of the module architecture of a thermal power generator set load forecasting device based on machine learning provided in one embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0034] In one embodiment, Figure 1 As shown, a load prediction device for a thermal power generator set based on machine learning is provided, which includes the following program modules: data acquisition, data preprocessing, feature engineering, model selection and training, model optimization and prediction, result output and feedback modules, wherein:

[0035] Data acquisition module: obtains load data, environmental data, fuel characteristics and equipment operation data from the thermal power generator control system, environmental monitoring system, fuel detection system and historical database;

[0036] Data preprocessing module: detect and clean outliers, fill missing values ​​and standardize the collected data;

[0037] Feature engineering module: construct multi-level lag features, dynamic weighted environmental features, autoencoder dimensionality reduction features, cross features and periodic features based on the characteristics of thermal power generation load;

[0038] Model selection and training module: Multi-model integration based on deep learning and machine learning, including LSTM, CNN-LSTM hybrid model, XGBoost, random forest and time series baseline model, and hyperparameter optimization through Bayesian optimization and reinforcement learning;

[0039] Model optimization and prediction module: Use dynamic weighting strategy to integrate multiple models, optimize the long-term performance of the model based on adaptive feedback mechanism, and perform real-time load prediction;

[0040] Result output and feedback module: outputs the load forecast results to the dispatching system, and uses the actual load data for model performance monitoring and dynamic adjustment through the feedback mechanism.

[0041] This embodiment overcomes the shortcomings of traditional load forecasting methods and effectively solves the main problems in the background technology through an end-to-end solution composed of modules such as comprehensive data collection, multi-level feature engineering, multi-model integration and adaptive optimization. The system collects multi-dimensional data from the control system, environmental monitoring system, fuel detection system and historical database of the thermal power generator set, including load data, environmental data, fuel characteristics and equipment operation data. This multi-dimensional data integration method makes the model not only rely on a single historical load data, but also combines multiple influencing factors such as environment, fuel, and equipment status, thereby overcoming the dependence of traditional methods on a single data source. Through the fusion of multi-dimensional data, the model can more comprehensively and accurately characterize the changing trend of the generator set load, effectively improve the accuracy of load forecasting, and make the prediction model more adaptable to different environmental conditions and equipment status.

[0042] This system constructs multi-level lag features, dynamically weighted environmental features, autoencoder dimensionality reduction features, cross features and periodic features through feature engineering modules. Traditional load forecasting usually only considers the historical value of the load, while this method greatly improves the input information quality of the model by extracting multiple features (such as lag features and periodic features) and generating cross features. In particular, the dynamic weighting of environmental features and the dimensionality reduction of the autoencoder can effectively avoid the interference of redundant features on the model while ensuring the integrity of key features. By enhancing the diversity and importance of input features, the model can more accurately identify short-term fluctuations and long-term trends of the load, making the prediction more robust and more adaptable, thus solving the problem of low prediction accuracy caused by the lack of feature engineering support in traditional methods.

[0043] Traditional load forecasting methods mostly rely on a single model, such as a simple linear regression or time series model, while this solution adopts a multi-model integration strategy. The integrated model includes LSTM, CNN-LSTM hybrid model, XGBoost, random forest and other algorithms. Through the complementary advantages of different algorithms, the overall prediction effect of the system is improved. At the same time, through the dynamic weighted fusion method, the model weight is dynamically adjusted according to the actual error to adapt to different load modes. Through multi-model integration and dynamic weighting, the system can effectively overcome the limitations of a single model, maintain high accuracy under different load modes, significantly improve the stability and scope of application of load forecasting, and solve the problem of low prediction accuracy of a single model in the background technology.

[0044] To enhance the real-time optimization capability of the model, this solution introduces reinforcement learning and Bayesian optimization in the model optimization and prediction module to achieve automatic tuning of hyperparameters and dynamically adjust the prediction error through real-time feedback. The reinforcement learning agent automatically adjusts the model parameters according to the actual prediction effect, and Bayesian optimization further improves the adaptability of the model. The combination of reinforcement learning and Bayesian optimization enables the model to be adaptively optimized according to real-time feedback, improves the accuracy and adaptability of the model under different working conditions, solves the problem that traditional models cannot be dynamically adjusted according to actual needs, and makes the prediction system more intelligent and flexible.

[0045] The result output and feedback module stores the prediction results through the time series database, and introduces a feedback mechanism to update the model according to the actual load conditions to ensure that the model can always adapt to the real-time changing load environment. The system compares the real-time prediction results with the actual data, and dynamically adjusts the model according to the error size to maintain the stability and accuracy of the prediction effect. Through real-time feedback and model updates, the system can continuously optimize the prediction effect, effectively solving the limitations of the traditional load forecasting system that lacks feedback and self-optimization. This enables the model to always reflect the load demand of the generator set in actual applications, improving the long-term adaptability of the system.

[0046] Through the coordinated application of the above technical modules, this system not only improves the problems of single data source, insufficient feature extraction, poor model adaptability, etc. in traditional load forecasting, but also significantly improves the prediction accuracy, adaptive ability and long-term stability of the model, thus providing an efficient, intelligent and stable technical solution for thermal power generation load forecasting. This systematic and innovative load forecasting method effectively overcomes the shortcomings of the background technology and has significant application value and innovation.

[0047] In this embodiment, the environmental data collected by the data collection module include air temperature, humidity, and wind speed, and the weight of each environmental feature is dynamically adjusted in combination with the actual load demand of the thermal power generating set.

[0048] This embodiment proposes a dynamic weighting strategy based on environmental characteristics, using the dynamic impact of environmental data such as temperature, humidity, and wind speed on load forecasting under different working conditions. The load of thermal power generating units is often significantly affected by environmental factors, especially in the case of large temperature differences or extreme climate changes, the impact of the environment on the load is particularly prominent. The dynamic weighted environmental characteristics in this embodiment innovatively improve the adaptability and accuracy of load forecasting by adjusting the weights of real-time environmental data. The specific effect analysis is as follows:

[0049] The dynamic weighting strategy of environmental characteristics enables the model to maintain high prediction accuracy under conditions of large environmental changes. For example, in hot weather, changes in temperature have a more significant impact on unit load; in this case, the method will automatically increase the weight of the temperature feature, making the model more sensitive to temperature fluctuations, thereby more accurately predicting the change trend of load. Through this adaptive feature weighting strategy, the model can flexibly respond to different weather conditions, greatly improving the adaptability of load forecasting, especially the reliability in extreme weather.

[0050] Thermal power generation load is affected by multiple environmental factors such as temperature, humidity, and wind speed. The impact of various environmental characteristics on the load varies at different times or in different geographical locations. The dynamic weighting strategy not only adjusts the weight of a single factor, but also comprehensively considers the synergy between environmental factors. For example, under conditions of low humidity and high wind speed, the impact of the environment on the load is more complex. The dynamic weighted multi-feature combination mode can better show the synergistic relationship of features under different environmental conditions, thereby improving the accuracy of the model under different weather conditions.

[0051] The dynamic weighting mechanism can also effectively reduce the interference of data noise. Since the sources of environmental feature data are diverse, there may be instability and accidental fluctuations. In traditional models, environmental noise may cause significant deviations in the predicted value. By dynamically adjusting the weight, the weight of the feature can be automatically reduced when the noise is large or the environmental data fluctuates violently, thereby reducing the impact of noise on load forecasting. This can ensure the high reliability and stability of the model while improving the robustness of the prediction.

[0052] Dynamically weighted environmental features enable the model to adaptively adjust the weights of input features under different time periods and operating conditions, improving the intelligence of the prediction system. Compared with static feature weights, dynamic weighting methods are more intelligent, have higher flexibility in real-time prediction, and can adapt to changing environmental conditions. This not only optimizes the prediction effect, but also simplifies the manual adjustment process of the model, making it more automated.

[0053] Through the dynamic weighting strategy of this embodiment, the model can adaptively adjust environmental factors in actual operation, significantly improving the accuracy, adaptability and robustness of load forecasting, especially under extreme weather and complex climate conditions.

[0054] In this embodiment, the multi-level lag features in the feature engineering module include intra-day lag features and inter-day lag features, which are used to capture short-term and long-term load change trends.

[0055] This embodiment captures the short-term fluctuations and long-term trends of the load by constructing multi-level lag features, providing the model with more comprehensive time-dependent information. Thermal power generation load has the characteristics of periodic fluctuations and accumulation, and the traditional single time lag feature is difficult to take into account both the short-term fluctuations and long-term trends of the load. The multi-level lag features of this embodiment include intra-day lag features (such as 1 hour, 3 hours, 6 hours, etc.) and inter-day lag features (such as the previous 1 day, the previous 7 days, etc.), which enhance the time memory ability of the model from different time scales.

[0056] By introducing intraday lag features, the model can capture short-term load fluctuations more precisely. For example, during peak hours when load demand fluctuates frequently, 1-hour and 3-hour lag features can reflect the short-term change trend of load, thus having a higher response speed in real-time forecasting. This makes the model more sensitive to changes in peak load and can track short-term fluctuations in real time, effectively avoiding the problem of inaccurate forecasting caused by time lag in traditional methods.

[0057] The daily lag feature records and analyzes the load data of the previous day, the previous 7 days, and other time points, so that the model can identify the long-term trend and seasonality of the load. For example, when the load gradually increases, the model can predict that the future load demand will show an upward trend and make adjustments in advance. The capture of long-term trends enables the model to not only adapt to short-term load fluctuations, but also has stronger long-term adaptability to meet the prediction needs in different time periods.

[0058] Multi-level lag features enable the model to obtain load change information from different time scales, enhancing the expressiveness of time series data. Compared with single lag features, multi-level lag features can better present the time series relationship in load data, so that the model not only relies on the most recent data when predicting, but also reviews and learns the long-term information in historical data, improving the global prediction ability of the model.

[0059] The multi-level hysteresis feature provides the model with multi-level information from short-term to long-term, making the model more robust to time series data. Especially in the case of alternating peak load and valley load, the multi-level hysteresis feature can better reflect the changing characteristics of the load, help the model maintain the stability of the forecast, and avoid the overall forecast deviation caused by abnormal data in a certain period of time.

[0060] Through the multi-level hysteresis feature construction of this embodiment, the model can take into account both short-term fluctuations and long-term trends in load forecasting, significantly improve the real-time, accuracy and stability of the forecast, and optimize the scope of application of the load forecasting system.

[0061] In this embodiment, the feature engineering module uses an autoencoder to perform dimensionality reduction processing on multi-dimensional features to reduce the impact of noise and retain the core structural information of the data;

[0062] The model selection and training module includes a network structure based on a mixture of convolutional neural networks and long short-term memory networks. The CNN layer extracts local features in the input sequence, and the LSTM layer is used to capture long-term dependencies.

[0063] Using autoencoders for feature dimensionality reduction effectively reduces redundant information and noise in the data, making the input features of the model more concise and efficient. In the load forecasting of thermal power generating units, the data sources are wide, the feature dimensions are many, there is a lot of noise and redundant information, and the traditional feature screening method is difficult to extract features comprehensively and efficiently. Autoencoders are an unsupervised learning model that can extract the main features of the data through dimensionality reduction compression and remove irrelevant information.

[0064] The autoencoder generates compact feature representations through training, reducing redundant information in high-dimensional data, thereby improving the representativeness of features. The feature dimension of thermal power generation load data is high. Through dimensionality reduction by the autoencoder, the model can only retain the most representative information for load forecasting, making the input features more effective. This process significantly reduces noise interference while retaining important feature information.

[0065] By reducing the dimension, the autoencoder generates a low-dimensional but high-information feature representation, reducing the feature dimension of the model input. This not only reduces the consumption of computing resources, but also speeds up the training of the model and improves the efficiency of the model in actual operation. Especially in the real-time prediction environment, the reduced-dimensional features can be input into the prediction model more quickly, achieving more efficient prediction.

[0066] There are a lot of noise and accidental fluctuations in thermal power load data. The autoencoder can filter out the noise in the data during the dimensionality reduction process and retain the main pattern of load changes. Compared with manual feature screening, the autoencoder dimensionality reduction can adaptively remove noise, effectively avoiding the errors that may be caused by traditional dimensionality reduction methods, making the prediction results more accurate and stable.

[0067] The autoencoder can extract the core structure of the data, so that the model has a higher generalization ability when processing data in different time periods and operating conditions. The features after dimensionality reduction not only have a higher amount of information, but also have stronger stability, which can avoid the overfitting problem caused by overly complex features. In this way, the model is more adaptable to different load conditions in practical applications.

[0068] In this embodiment, the model selection and training module adopts a dynamic weighted multi-model integration strategy to dynamically adjust the weight of the model according to real-time error feedback to improve the overall prediction accuracy of the model.

[0069] In this embodiment, the model optimization and prediction module optimizes the hyperparameter settings of the model through a reinforcement learning algorithm, wherein the reinforcement learning agent automatically adjusts the hyperparameters of the model based on prediction error feedback.

[0070] In this embodiment, the model optimization and prediction module includes a GAN-based adversarial training mechanism, the generator is used to output a load prediction value, and the discriminator feeds back an error signal to correct the prediction result of the generator, thereby reducing the error of load prediction.

[0071] In this embodiment, the model optimization and prediction module introduces a self-attention mechanism into the LSTM model, so that the model can dynamically allocate the importance of features at each time step, thereby improving the accuracy of load prediction.

[0072] In this embodiment, the model optimization and prediction module performs load prediction based on the Transformer structure, which is used to capture long-term periodic relationships in load data and is suitable for predicting long-term load change trends.

[0073] In this embodiment, the result output and feedback module stores the predicted values ​​in a time series database, and dynamically updates the prediction model in combination with the actual load data to maintain the adaptability of the model to different operating environments.

[0074] This embodiment proposes a hybrid network structure that combines a convolutional neural network (CNN) with a long short-term memory network (LSTM). The CNN layer extracts local features of the data, and the LSTM layer processes the long-term dependencies of the time series. The load data of thermal power generating units has multi-dimensional complexity and time dependency. Traditional single models often find it difficult to take into account both the local features and long-term dependencies of the data. The convolution-LSTM hybrid network structurally combines the advantages of the two models, thereby improving the accuracy and applicability of load forecasting.

[0075] The CNN layer extracts local features of the data through convolution operations and can effectively identify local patterns in the input data. In thermal power load data, local features often reflect detailed information about short-term load changes. The addition of the CNN layer enables the model to identify and utilize subtle features in the load data, enhancing the model's ability to capture short-term load changes. Compared with traditional time series models, the CNN layer can more finely analyze the local features of the data, allowing the model to show higher accuracy when processing complex and variable load data.

[0076] The LSTM layer has the ability to handle long-term dependencies and is suitable for periodicity and long-term trends in load data. After the local features extracted by the convolution layer are processed by the LSTM layer, the model can make full use of the time series information of the load data to accurately capture the long-term load trend. For example, the gradually accumulated periodic changes in the load data can be effectively learned and stored through LSTM, and future load trends can be predicted. Compared with traditional models, the LSTM layer can retain important state information in the time series, making the prediction results more continuous and stable.

[0077] The convolution-LSTM hybrid network improves the model's adaptability to complex load patterns through multi-level feature fusion. CNN is responsible for extracting local features, and LSTM is responsible for capturing time dependencies. The combination of the two forms a multi-dimensional feature expression, which enables the model to maintain sensitivity in short-term predictions and robustness in long-term predictions. This multi-level feature fusion capability enables the model to adapt to the variability and complexity of thermal power generation loads, effectively improving the adaptability of load forecasting.

[0078] The convolution layer significantly reduces the computational complexity of the model through the parameter sharing mechanism of the convolution kernel, while retaining the important features in the data. The convolution operation allows the model to efficiently express data features in a lower dimension, while LSTM further improves the time series modeling capability, optimizing the overall model's prediction efficiency. Compared with a single LSTM network, the convolution-LSTM hybrid network improves the prediction speed while maintaining high accuracy, making it suitable for real-time prediction applications.

[0079] The convolution-LSTM hybrid network of this embodiment achieves accurate modeling of thermal power load data through the combination of local features and long-term dependencies, improves the short-term sensitivity and long-term adaptability of the model, and provides an efficient technical solution for complex load forecasting.

[0080] This embodiment proposes a dynamic weighted multi-model integration strategy, which uses real-time error feedback to adjust the weight of each model to achieve adaptive prediction of different load modes. Traditional load forecasting methods usually rely on a single model, which is difficult to fully adapt to the changing laws of thermal power generation load in different time periods and operating conditions. The dynamic weighted multi-model integration strategy improves the overall stability and adaptability of the prediction by combining the advantages of multiple models.

[0081] The integrated model includes algorithms such as LSTM, CNN-LSTM, XGBoost and random forest. Each model has its own strengths under different load characteristics. Through multi-model fusion, the advantages of different algorithms can be effectively integrated to make the overall prediction effect more accurate. For example, LSTM has advantages in time series modeling, while XGBoost performs well in processing nonlinear features. The dynamic weighting mechanism adjusts the weight of the model according to the actual error situation, and automatically selects the most appropriate model combination under different load modes to achieve the best prediction effect.

[0082] The dynamic weighting strategy is based on a real-time error feedback mechanism, which can automatically adjust the weight of the model according to the size and change trend of the current forecast error. For example, when the error of a certain model increases, the system automatically reduces the weight of the model and increases the weight of other models to minimize the overall forecast error. In this way, the system can flexibly adapt under different load conditions, improve the adaptability of the forecast system, and is particularly suitable for situations where load fluctuations are frequent and uncertainty is high.

[0083] A single model may be unstable when facing sudden load changes, while the dynamic weighted multi-model integration strategy improves the long-term stability of load forecasting through the collaboration of multiple models. The dynamic weighting mechanism can adjust the weights of different models in real time during periods of large load fluctuations, thereby effectively avoiding the overall forecast error caused by the failure of a single model. The synergy of multiple models improves the robustness of the forecast, allowing the system to maintain a high degree of stability when facing various complex load patterns.

[0084] Traditional load forecasting systems often require manual selection and adjustment of models, while the dynamic weighted integration strategy automates this process. The automatic adjustment mechanism of weights makes the model adjustment process more intelligent, reduces the need for manual intervention, and improves the automation level of the system. This adaptive dynamic weighting strategy optimizes the forecasting process and realizes an intelligent and automated forecasting method.

[0085] Through the dynamic weighted multi-model integration strategy of this embodiment, the system can flexibly adjust the model weights according to real-time error feedback, effectively integrate the advantages of different algorithms, improve the accuracy, stability and adaptability of load forecasting, and significantly optimize the overall effect of load forecasting.

[0086] This embodiment proposes a reinforcement learning optimization strategy, which optimizes the model's hyperparameter settings in real time through a reinforcement learning algorithm to achieve adaptive tuning of the model. The changes in thermal power generation load are highly complex and unpredictable. The parameters of traditional load forecasting systems are mostly fixed, making it difficult to adapt to the parameter adjustment requirements in complex environments. The reinforcement learning optimization strategy automatically adjusts the model hyperparameters by learning error feedback signals, thereby improving the flexibility and intelligence level of the load forecasting system.

[0087] The reinforcement learning algorithm can dynamically optimize the model's hyperparameters by continuously learning error feedback during the real-time prediction process. For example, the model can adjust key parameters such as the learning rate and regularization coefficient according to the size of the current prediction error to adapt to the prediction needs under different load conditions. Compared with models with fixed parameters, the reinforcement learning optimization strategy can achieve real-time adjustments based on actual working conditions, improving the model's adaptability to sudden load changes.

[0088] The hyperparameter setting of traditional models usually requires a lot of manual debugging, which is time-consuming and labor-intensive with inconsistent results. Through reinforcement learning, the system can learn the optimal parameter combination in error feedback, thereby achieving automatic parameter adjustment. This adaptive tuning mechanism enables the system to autonomously optimize parameters during operation, reducing dependence on manual parameter adjustment, improving the intelligence level of the system, and facilitating large-scale practical applications.

[0089] The reinforcement learning agent continuously adjusts parameters based on error feedback, enabling the model to self-correct and improve in long-term forecasting. As the system continuously optimizes the hyperparameter combination, the prediction error gradually decreases, allowing the model to maintain high accuracy and stability in long-term operation. Compared with models with static parameters, the reinforcement learning optimization strategy can adapt to different load conditions and environmental changes, improving the long-term stability of the load forecasting system.

[0090] In one embodiment, in thermal power generation load forecasting, load data is often sudden and volatile, especially under extreme conditions, and traditional models are difficult to cope with such complex changes. Through GAN adversarial training, the model can not only handle conventional load fluctuations, but also cope with sudden load changes through continuous optimization of the generator. While the discriminator continuously identifies the difference between the generator output and the actual data, it can guide the generator to gradually learn complex load change patterns, improve the robustness of the model in complex scenarios, and ensure that the system can still make stable predictions under abnormal conditions.

[0091] Traditional load forecasting models are usually highly dependent on historical data during training, making it difficult to quickly respond to sudden load changes in real-time operation. The adversarial mechanism of the generative adversarial network allows the model to learn and adjust through real-time feedback, reducing the reliance on large amounts of historical data. While the generator optimizes the prediction results through the adversarial training process, the discriminator can promptly detect deviations in the generated results and provide feedback. This real-time iterative training mechanism enables the model to have stronger real-time adaptability.

[0092] The generative adversarial network of this embodiment realizes automatic tuning and learning of the model through the automated competition process between the generator and the discriminator, reducing the need for human intervention. The generator continuously tries to generate data that is closer to the actual load, while the discriminator automatically detects the differences in the generated data, and the entire process does not require human intervention. Compared with traditional models that require manual parameter setting and algorithm adjustment, the GAN model has a high degree of automation and can autonomously learn load change patterns and optimize prediction performance. Since the generative adversarial network can continuously optimize the output of the generator through long-term adversarial training, the model has the ability to self-correct in long-term operation. This long-term adaptability enables the model to not only cope with short-term load fluctuations, but also continuously improve the prediction accuracy in long-term operation as the environment and equipment status change, thereby maintaining the stability and reliability of the prediction system.

[0093] Through the generative adversarial network adversarial training mechanism of this embodiment, the load forecasting system has the ability of adaptive error correction when dealing with complex load patterns and sudden fluctuations, which significantly improves the prediction accuracy, robustness and long-term adaptability, while reducing dependence on historical data and realizing a highly automated prediction correction process.

[0094] This embodiment proposes a long short-term memory network (LSTM) that introduces a self-attention mechanism, which enables the model to dynamically identify and weight the importance of different time steps and features, thereby improving the accuracy and interpretability of load forecasting. The self-attention mechanism can dynamically assign weights to each input feature when the model processes multidimensional features and long time series, helping the model focus on the time steps or features that are most important for the current prediction, thereby achieving more accurate predictions.

[0095] When processing long time series, traditional LSTM networks treat the features of each time step with equal importance and cannot distinguish the degree of influence of different time steps on the current prediction. After introducing the self-attention mechanism, the model can dynamically assign weights to different time steps according to the specific changes in load data, so that the model focuses on historical time points that are more meaningful to the current prediction. For example, in the case of large load fluctuations, the model can automatically adjust the weights, pay more attention to recent data, reduce the impact of long-term data, and thus improve prediction accuracy.

[0096] Thermal power generation load forecasting often involves multi-dimensional input data, such as ambient temperature, humidity, coal quality, equipment operating status, etc. The self-attention mechanism enables the model to dynamically judge and weight the importance of different input features, avoiding feature redundancy or over-reliance on a certain feature. For example, when the temperature changes significantly at a certain moment, the model will automatically give a higher weight to the temperature feature, making the prediction result more accurate. Compared with the traditional model that treats all features equally, the self-attention mechanism can significantly improve the model's feature expression ability. The self-attention mechanism can effectively focus on the key time points of load changes in load forecasting. For example, if the load changes drastically at a certain moment, the model will automatically increase the attention to that moment and its adjacent time points, ignoring those time steps that contribute less to the prediction. This ensures that the model can more accurately capture key change points when dealing with complex load fluctuations, and improve the sensitivity and accuracy of short-term load forecasting.

[0097] In one embodiment, since the self-attention mechanism can explicitly display the weight distribution of each time step and feature, the decision-making process of the model becomes more transparent. For the load forecasting system, dispatchers can understand the key factors behind the load changes by observing the attention distribution of the model, so as to make more reasonable scheduling decisions. Compared with the traditional black box model, the self-attention mechanism gives the model stronger interpretability, helping users understand the logical basis of the prediction results.

[0098] When processing long time series data, LSTM networks are prone to information overload or forgetting problems, resulting in poor prediction results. The self-attention mechanism can effectively reduce the pressure of long sequence input on the model by weighting key time steps, avoiding information loss or excessive memory problems in the model in long-term dependencies, thereby improving the stability and accuracy of the model when processing long time series.

[0099] By introducing the self-attention mechanism in this embodiment, the load forecasting system not only improves the processing capabilities of long time series and multi-dimensional data, but also enhances the sensitivity to key time steps and important features, significantly improves the accuracy, stability and interpretability of the model, and makes load forecasting more intelligent and operational.

[0100] In this embodiment, a load forecasting model based on the Transformer structure is proposed, which can effectively capture the long-term periodic relationship in the load data. The Transformer structure has achieved remarkable success in fields such as natural language processing. Its superior parallel processing capability and long-distance dependency capture capability enable it to perform well in load forecasting, especially for processing complex time series data. Traditional LSTM networks have the problem of information forgetting when processing long time series, while Transformer can effectively capture the relationship between each time step in the sequence through the self-attention mechanism. Transformer does not rely on the order of time steps, but considers the data of all time steps at the same time through the attention mechanism, so that the model can fully obtain the information in the long time series. This method can more accurately identify long-term trends and periodic changes when dealing with complex load fluctuations, thereby improving the accuracy of load forecasting.

[0101] The Transformer structure has powerful parallel computing capabilities. Compared with traditional sequence models (such as LSTM) that need to process time series data step by step, the Transformer can process the entire sequence simultaneously. This parallel computing feature significantly speeds up the training of the model, especially on large-scale data sets, which can effectively reduce the training time. Improving training efficiency means that the model can perform multiple iterations of optimization in a shorter time, making it more adaptable to rapidly changing load environments in practical applications. The self-attention mechanism of the Transformer model allows the model to focus on important features in different time periods, especially when processing periodic load data, it can accurately capture the characteristic performance of each period. By dynamically adjusting the attention weight, the model can learn the periodic variation law of load data and further optimize the accuracy of long-term prediction. This sensitivity to periodic features makes the model more adaptable in thermal power generation load forecasting. The self-attention mechanism of the Transformer model not only improves the accuracy of load forecasting, but also enhances the interpretability of the model. During the training process, the model can explicitly show the degree of influence of each input feature on the final prediction result. By analyzing the attention weight, dispatchers can understand which features play a key role in the prediction, making the decision-making process of the model more transparent. This interpretability provides a more reliable basis for load forecasting and helps managers make more reasonable decisions in actual scheduling. The introduction of the Transformer structure enables the model to maintain good generalization capabilities when facing different load patterns. Through the self-attention mechanism, the model can flexibly adjust the features and time points of interest, reduce dependence on historical data, and enhance adaptability to new data and emergencies. In a complex and dynamic power market environment, Transformer can effectively cope with the variability of loads and maintain good forecasting performance.

[0102] In summary, the Transformer-based load forecasting model of this embodiment improves the accuracy of load forecasting, training efficiency and model interpretability by effectively capturing long-term periodic relationships, making it an important technological innovation in the field of thermal power generation.

[0103] The specific implementation content of each module can be found in the above definition of the load forecasting system for thermal power generators based on machine learning, which will not be repeated here.

[0104] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A load forecasting system for thermal power generating units based on machine learning, including data acquisition, data preprocessing, feature engineering, model selection and training, model optimization and prediction, result output and feedback modules, characterized in that: Data acquisition module: obtains load data, environmental data, fuel characteristics and equipment operation data from the thermal power generator control system, environmental monitoring system, fuel detection system and historical database; Data preprocessing module: detect and clean outliers, fill missing values ​​and standardize the collected data; Feature engineering module: construct multi-level lag features, dynamic weighted environmental features, autoencoder dimensionality reduction features, cross features and periodic features based on the characteristics of thermal power generation load; Model selection and training module: Multi-model integration based on deep learning and machine learning, including LSTM, CNN-LSTM hybrid model, XGBoost, random forest and time series baseline model, and hyperparameter optimization through Bayesian optimization and reinforcement learning; Model optimization and prediction module: Use dynamic weighting strategy to integrate multiple models, optimize the long-term performance of the model based on adaptive feedback mechanism, and perform real-time load prediction; Result output and feedback module: outputs the load forecast results to the dispatching system, and uses the actual load data for model performance monitoring and dynamic adjustment through the feedback mechanism.

2. The load forecasting system for thermal power generating units based on machine learning according to claim 1 is characterized in that: The environmental data collected by the data collection module include temperature, humidity, and wind speed, and the weight of each environmental feature is dynamically adjusted in combination with the actual demand of the thermal power generator set load.

3. The load forecasting system for thermal power generating units based on machine learning according to claim 1 is characterized in that: The multi-level lag features in the feature engineering module include intra-day lag features and inter-day lag features, which are used to capture short-term and long-term load change trends.

4. The load forecasting system for thermal power generating units based on machine learning according to claim 1, characterized in that: The feature engineering module uses an autoencoder to reduce the dimensionality of multi-dimensional features to reduce the impact of noise and retain the core structural information of the data; The model selection and training module includes a network structure based on a mixture of convolutional neural networks and long short-term memory networks. The CNN layer extracts local features in the input sequence, and the LSTM layer is used to capture long-term dependencies.

5. The load forecasting system for thermal power generating units based on machine learning according to claim 1, characterized in that: The model selection and training module adopts a dynamic weighted multi-model integration strategy to dynamically adjust the model weights according to real-time error feedback to improve the overall prediction accuracy of the model.

6. The machine learning-based thermal power generation unit load forecasting system according to claim 1, characterized in that: The model optimization and prediction module optimizes the hyperparameter settings of the model through a reinforcement learning algorithm, wherein the reinforcement learning agent automatically adjusts the hyperparameters of the model based on the prediction error feedback.

7. The load forecasting system for thermal power generating units based on machine learning according to claim 1, characterized in that: The model optimization and prediction module includes a GAN-based adversarial training mechanism, in which the generator is used to output a load prediction value, and the discriminator feeds back an error signal to correct the prediction result of the generator, thereby reducing the error of load prediction.

8. The machine learning-based thermal power generation unit load forecasting system according to claim 1, characterized in that: The model optimization and prediction module introduces a self-attention mechanism into the LSTM model, so that the model can dynamically allocate the importance of features at each time step, thereby improving the accuracy of load prediction.

9. The machine learning-based thermal power generation unit load forecasting system according to claim 1, characterized in that: The model optimization and prediction module performs load prediction based on the Transformer structure, which is used to capture long-term periodic relationships in load data and is suitable for predicting long-term load change trends.

10. The load forecasting system for thermal power generating units based on machine learning according to claim 1, characterized in that: The result output and feedback module stores the predicted values ​​in a time series database, and dynamically updates the prediction model in combination with the actual load data to maintain the adaptability of the model to different operating environments.

Images