Generation power load prediction scheduling system based on machine learning

Abstract

The invention discloses a power generation power load prediction scheduling system based on machine learning, and relates to the technical field of power scheduling, and the system comprises a dynamic perception control module which receives external environment data and internal prediction performance indexes in real time, and generates a data fusion control instruction and a physical guidance control instruction; the hybrid prediction execution module outputs a power load prediction curve and a stability risk index through physical mechanism simulation of the physical guide unit, multi-source data weighted fusion of the dynamic fusion unit and a neural network model of the time sequence prediction unit; the online self-adaptive fine tuning module monitors prediction errors and dynamically optimizes strategy network parameters; according to the method, a data driving method and physical mechanism knowledge are deeply fused, the problems of multi-source data value dynamic change and extreme scene data shortage are effectively solved, the prediction reliability and physical consistency are improved, and meanwhile, organic linkage of prediction and scheduling is achieved.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, specifically to a power generation load prediction and dispatching system based on machine learning. Background Technology

[0002] With the deepening of energy transition and the integration of a high proportion of renewable energy, the operational complexity of power systems has increased significantly, placing higher demands on the accuracy and robustness of load forecasting. Accurate load forecasting is the foundation for formulating economical and safe power generation plans and dispatching strategies. In existing technologies, load forecasting systems typically rely on historical data to drive models and attempt to integrate multi-source data, such as meteorological and economic data, to improve forecasting performance. For example, prior art publication number CN116544934B discloses a power dispatching method and system based on power load forecasting, which focuses on selecting a recommended scheme by generating and evaluating the fitness of multiple power supply schemes after obtaining the predicted load, thereby improving the precision of dispatching. Another prior art publication number CN114595861A involves a medium- and long-term power load forecasting method based on MSTL and LSTM models, which focuses on using attention mechanisms to fuse data features from different sources. These technologies represent two mainstream improvement directions: one is to optimize the downstream dispatching application of forecast results, and the other is to improve the data fusion method at the front end of the forecasting model.

[0003] However, the aforementioned existing technologies still have a limitation when dealing with complex and ever-changing real-world scenarios: existing systems struggle to effectively address the complex issues of dynamically changing value of multi-source heterogeneous data and the scarcity of historical data for extreme / emerging scenarios. Specifically, under normal operating conditions, data fusion methods driven by historical data and combined with fixed weights are still applicable; however, when encountering complex weather conditions or special power grid operating modes, the reliability and value of different data sources change drastically with the scenario, and the system lacks a mechanism for intelligently sensing and evaluating this. Furthermore, for such extreme or emerging scenarios lacking historical samples, purely data-driven models cannot learn sufficient patterns, leading to increased prediction bias, and lack the ability to integrate knowledge of the physical laws of the power system to compensate for information blind spots. The consequence is that prediction results may deviate from physical reality, making subsequent, even the most refined scheduling optimizations based on unreliable predictions, ultimately affecting the accuracy of overall decision-making. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a power generation load forecasting and scheduling system based on machine learning. This system can break through the paradigm of static data fusion and pure data-driven approaches, intelligently perceive changes in the external environment and internal state, and dynamically integrate data-driven methods with physical mechanism knowledge. In this way, it can generate more reliable load forecasts that conform to physical laws when facing complex and unknown scenarios.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a power generation load forecasting and dispatching system based on machine learning, comprising: The dynamic perception control module is used to receive external environmental data and internal prediction performance indicators in real time, and generate data fusion control instructions and physical guidance control instructions based on the external environmental data and internal prediction performance indicators. The external environmental data includes at least meteorological warning signals, and the internal prediction performance indicators include at least short-term prediction error statistics. A hybrid prediction execution module, communicatively connected to the dynamic sensing control module, is used to receive the data fusion control command and the physical guidance control command, and to perform load prediction. The hybrid prediction execution module includes: The physical guidance unit is used to activate and run the power system physical mechanism simulation model according to the physical guidance control command when the preset conditions are met, and generate a physical simulation feature sequence. The dynamic fusion unit is used to perform weighted fusion of the input multi-source historical load characteristic data according to the data fusion control command, and to splice the fusion result with the physical simulation characteristic sequence to form the final input feature; The time-series prediction unit is used to receive the final input features and output the power load prediction curve for a specified future period through a trained time-series prediction neural network model. The dynamic perception control module and the hybrid prediction execution module work together to form an adaptive learning prediction engine. By dynamically adjusting the data fusion control command and the physical guidance control command, it addresses the complex issues of dynamic changes in the value of multi-source heterogeneous data and the scarcity of historical data in extreme new scenarios.

[0006] Furthermore, the dynamic perception control module includes a policy network, which uses state vectors... As input, the output includes a data fusion weight vector. With physical guiding strength coefficient The action command, the state vector The construction of the formula is: ,in, This represents a scalar value indicating the coded weather warning level. This represents a scalar quantization of uncertainty calculated based on the prediction errors of the most recent N periods. Indicates the scalar value for holidays. Represents the scalar value at the current moment.

[0007] Furthermore, the physical guiding unit in the physical guiding strength coefficient Greater than the preset activation threshold It is activated at that time; After activation, the physical guidance unit, based on the current power grid topology parameters and basic load data, calls the power system power flow calculation simulation program to simulate and generate theoretical load distribution curves under at least one extreme scenario, which serve as the physical simulation feature sequence. The physical simulation feature sequence is multiplied by the physical guidance intensity coefficient before being concatenated with the output of the dynamic fusion unit. Weighting is applied.

[0008] Furthermore, the multi-source historical load characteristic data received by the dynamic fusion unit includes at least the encoding characteristics of historical load time series data, the encoding characteristics of refined grid meteorological historical data, and the encoding characteristics of macroeconomic operation data; The data fusion control instruction is specifically as follows: for the weight allocation vector of the encoded feature channels of different data sources, the dynamic fusion unit realizes the weighted fusion of feature channels by multiplying the weight allocation vector with the corresponding feature channels element by element.

[0009] Furthermore, the time-series prediction unit employs a time-series prediction neural network model that is a long short-term memory network, a gated recurrent unit, or a time-series fusion Transformer model. The hybrid prediction execution module is optimized by combining offline joint training with online incremental fine-tuning. The offline joint training aims to minimize the overall loss of load prediction and simultaneously optimizes the policy network parameters of the dynamic perception control module and the neural network model parameters of the time series prediction unit.

[0010] Furthermore, the offline joint training is achieved through a two-layer optimization, with the outer layer optimizing the objective function. With inner optimization objective function The following relationship must be satisfied: Inner fixed strategy network parameters Optimize the parameters of the time series prediction model for: ; The outer layer optimizes the network parameters based on the results of the inner layer. for: ; in, Represents the validation dataset. To predict the model's loss on the training set, To predict the model's loss on the validation set, This indicates that inner optimization depends on The optimal prediction model parameters.

[0011] Furthermore, the system also includes an online adaptive fine-tuning module; The online adaptive fine-tuning module continuously monitors the prediction error output by the hybrid prediction execution module. When it detects that the prediction error in a specific scenario type continuously exceeds a set threshold, it triggers online fine-tuning of the policy network in the dynamic perception control module. The online fine-tuning uses newly generated scene data to adjust the policy network parameters through a finite number of gradient updates in order to quickly adapt to the specific scene.

[0012] Furthermore, the power load forecast curve output by the hybrid forecast execution module is further input to the scheduling optimization module; The scheduling optimization module is used to generate generator start-up and shutdown plans and output plans based on the power load forecast curve and in combination with preset grid operation constraints and generation cost functions.

[0013] Furthermore, when generating control commands, the dynamic perception control module also uses internal prediction performance indicators to determine the deviation between the peak and valley characteristics of the prediction curve and the average peak and valley characteristics of the same period in history. This allows the generated commands to guide the hybrid prediction execution module to produce a load prediction curve that is more favorable for subsequent scheduling decisions.

[0014] Furthermore, the power system physical mechanism simulation model that runs when the physical guidance unit is activated is a power grid simulation model based on digital twins; The digital twin power grid simulation model receives the scenario parameters contained in the physical guidance and control command, performs parallel Monte Carlo simulation, generates the physical simulation feature sequence, and outputs the stability risk index of the power grid operation under the scenario. The stability risk index is output together with the power load forecast curve.

[0015] Compared with existing technologies, this machine learning-based power generation load forecasting and dispatching system has the following advantages: I. This invention establishes a dynamic perception control module to receive external environmental data and internal predictive performance indicators in real time, constructs a state vector, and generates dynamic data fusion control commands and physical guidance control commands. Combined with the power system physical mechanism simulation model of the physical guidance unit in the hybrid prediction execution module and the weighted fusion of multi-source historical load characteristic data of the dynamic fusion unit, this invention deeply integrates data-driven methods with physical mechanism knowledge. It can effectively address the complex issues of dynamic changes in the value of multi-source heterogeneous data and the scarcity of historical data in extreme or emerging scenarios. This avoids prediction bias caused by the lack of historical samples in extreme or emerging scenarios due to the pure data-driven model, prevents prediction results from deviating from physical reality, thereby improving the reliability and physical consistency of load prediction and providing a solid foundation for subsequent scheduling decisions.

[0016] Second, this invention employs a model optimization mechanism that combines offline stage joint training with online stage incremental fine-tuning. This mechanism simultaneously optimizes the strategy network parameters of the dynamic perception control module and the neural network model parameters of the time-series prediction unit. Furthermore, it utilizes an online adaptive fine-tuning module to continuously monitor prediction errors and dynamically adjust the strategy network parameters when prediction accuracy is insufficient in specific scenarios. This allows for rapid adaptation to different scenario characteristics, continuously improving the system's prediction accuracy and generalization ability. Additionally, the scheduling optimization module combines load prediction curves with grid operation constraints and generation cost functions to generate generator start-up and shutdown plans and output plans, achieving an organic connection between load prediction and scheduling optimization. This enhances the economy and security of power system operation and reduces scheduling decision risks.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a schematic diagram of the overall system architecture and data flow of the present invention; Figure 2 This is a schematic diagram of the internal structure of the hybrid prediction execution module of the present invention; Figure 3 This is a schematic diagram of the offline joint training and online adaptive fine-tuning process of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1 like Figures 1 to 3As shown, this embodiment provides a power generation load forecasting and scheduling system based on machine learning. This system dynamically senses changes in the external environment and internal forecasting status, and dynamically integrates data-driven methods and physical mechanism knowledge to effectively address the complex problems of dynamic changes in the value of multi-source heterogeneous data and the scarcity of historical data in extreme emerging scenarios, thereby improving the reliability and physical consistency of load forecasting.

[0022] In this embodiment, the system mainly includes a dynamic perception control module, a hybrid prediction execution module, an online adaptive fine-tuning module, and a scheduling optimization module. The dynamic perception control module and the hybrid prediction execution module work together to form an adaptive learning prediction engine. The online adaptive fine-tuning module is used to optimize the parameters of the dynamic perception control module in real time. The scheduling optimization module generates a scheduling plan based on the prediction results. All modules are connected through a data bus to ensure the real-time performance and stability of data transmission.

[0023] In this embodiment, the core function of the dynamic perception control module is to receive external environmental data and internal predictive performance indicators in real time, and generate data fusion control commands and physical guidance control commands. Its core component is the policy network.

[0024] In this embodiment, the external environmental data mainly includes meteorological warning signals, which are pushed in real time by the meteorological department's monitoring system, covering various meteorological disaster warning information such as rainstorms, typhoons, cold waves, and high temperatures. Internal prediction performance indicators include short-term prediction error statistics and the deviation between the peak-valley characteristics of the prediction curve and the historical average peak-valley characteristics for the same period. The short-term prediction error statistics are obtained by calculating the deviation between the predicted load value and the actual load value in the most recent prediction period. The deviation is obtained by comparing the peak-valley difference, peak value, and valley value of the current prediction curve with the corresponding parameters of the historical average peak-valley characteristics for the same period.

[0025] All received data must undergo preprocessing, including data cleaning, normalization, and encoding conversion, to ensure that the data format is uniform, the range is reasonable, and the input requirements of the policy network are met. For example, weather warning signals are encoded into weather warning level scalars, prediction errors are statistically calculated to obtain uncertainty quantification scalars, holiday identifier scalars are generated based on the date, and the current time scalar is obtained based on the system clock.

[0026] In this embodiment, the policy network uses state vectors. As input, the state vector The construction formula is: in This is a scalar representation of the coded weather warning level. The coding rules are set based on the severity of the weather warning; the higher the warning level, the higher the scalar representation. The larger the value, the higher the range from 0 to 1, for example, when there is no warning. A value of 0 indicates a typical warning situation. The value is 0.3, indicating a severe warning. The value is 0.7, especially during severe warnings. It is 1.0; This is a scalar quantification of uncertainty calculated based on the prediction errors of the most recent N periods. N is set according to the prediction period type; for hourly predictions, N is typically 24, and for daily predictions, N is typically 7. It is calculated as the ratio of the standard deviation of the prediction errors of the most recent N periods to the average load, with a value ranging from 0 to 1. The larger the value, the higher the uncertainty of near-term forecasts; For holidays, use scalars; for weekdays... It is 0, on weekends and public holidays. The value is 1, used to distinguish the impact of different date types on the load; As a scalar for the current time, a day of 24 hours is divided into several time periods, each time period corresponds to a fixed value, with a value range from 0 to 1. For example, the early morning time period corresponds to 0.1, the morning peak time period corresponds to 0.4, the evening peak time period corresponds to 0.8, and the late night time period corresponds to 0.9, which is used to reflect the intraday time-series variation characteristics of the load.

[0027] Policy Network Structure and Output: The policy network adopts a fully connected neural network structure, including an input layer, hidden layers, and an output layer. Input layer dimensions and state vector. The dimensions are consistent and set to 4-dimensional; there are 2 hidden layers, and the number of neurons in each layer is adjusted according to the actual training effect, usually 32 to 128. The hidden layers use the ReLU activation function to enhance the non-linear fitting ability of the network; the output layer uses the sigmoid activation function to map the output value to between 0 and 1 to ensure that the output control command parameters are reasonable.

[0028] The output of the policy network is an action instruction, which includes a data fusion weight vector. With physical guiding strength coefficient The data fusion weight vector The dimension is consistent with the number of multi-source feature channels in the dynamic fusion unit, which is 3-dimensional in this embodiment, corresponding to the weights of historical load time-series data coding features, refined grid meteorological historical data coding features, and macroeconomic operation data coding features, respectively; physical guidance intensity coefficient It is a single scalar used to control the activation state and intensity of the physical guiding unit.

[0029] In this embodiment, the hybrid prediction execution module receives control commands sent by the dynamic perception control module and performs load prediction. It includes a physical guidance unit, a dynamic fusion unit, and a timing prediction unit.

[0030] In this embodiment, the core function of the physical guidance unit is to activate when preset conditions are met, run the power system physical mechanism simulation model, and generate a physical simulation feature sequence.

[0031] In this embodiment, the physical guidance unit receives the physical guidance intensity coefficient in real time. and with the preset activation threshold Compare. Preset activation threshold. The value was calibrated using offline verification data, with a range of 0.3 to 0.7; in this embodiment, it is set to 0.5. When When, the physical guidance unit is activated; when When the physical guidance unit is not activated, the physical simulation feature sequence is a zero vector and does not participate in subsequent feature splicing.

[0032] In this embodiment, after the physical guidance unit is activated, it calls a digital twin-based power grid simulation model as the simulation model of the power system's physical mechanism. This digital twin power grid simulation model is constructed based on the topology, equipment parameters, and line parameters of the actual power grid, and has a high degree of consistency with the actual power grid, enabling it to accurately simulate the power grid's operating state.

[0033] First, the digital twin power grid simulation model receives the current power grid topology parameters and basic load data. The power grid topology parameters include bus connection relationships, transformer parameters, line impedance parameters, etc., and the basic load data is the actual load data at the current moment after preprocessing. Then, based on the scenario parameters in the physical guidance control command, the type of extreme scenario to be simulated is determined. Extreme scenarios include extreme weather scenarios, power grid fault scenarios, peak electricity consumption scenarios, etc. In this embodiment, at least three typical extreme scenarios are simulated. Finally, parallel Monte Carlo simulation is started. Through the power system power flow calculation simulation program, the power flow distribution under each extreme scenario is solved, and the corresponding theoretical load distribution curve is generated. Multiple theoretical load distribution curves constitute the physical simulation feature sequence.

[0034] In this embodiment, the physical simulation feature sequence needs to be multiplied by the physical guidance intensity coefficient before being concatenated with the output of the dynamic fusion unit. Weighting is applied. The purpose of weighting is to dynamically adjust the influence of physical simulation features on the final prediction result based on the physical guidance intensity coefficient. The larger the value, the higher the influence weight of the physical simulation features, thereby strengthening the guiding role of physical mechanisms in prediction under extreme scenarios and ensuring that the prediction results conform to physical laws.

[0035] In this embodiment, the core function of the dynamic fusion unit is to perform weighted fusion of multi-source historical load characteristic data according to the data fusion control command to form fused characteristics.

[0036] In this embodiment, the multi-source historical load characteristic data received by the dynamic fusion unit includes historical load time series data coding characteristics, refined grid meteorological historical data coding characteristics, and macroeconomic operation data coding characteristics.

[0037] Historical load time-series data consists of load monitoring data from the past few years, stored in hourly or daily units. After time-series data encoding, a fixed-dimensional feature vector is obtained. The encoding method uses a temporal convolutional network to extract time-series features, capturing the trend, periodicity, and randomness of the load. Refined gridded meteorological historical data consists of gridded meteorological data covering the power grid supply area, including indicators such as temperature, humidity, wind speed, and precipitation. After being divided into grids, the meteorological features of each grid are extracted, and then spatial pooling is used to obtain the overall meteorological feature vector of the region. Macroeconomic operation data includes regional GDP growth rate, industrial output, the proportion of the tertiary industry, population, and other economic and social indicators. After normalization and encoding, an economic feature vector is obtained.

[0038] In this embodiment, the data fusion control command is a data fusion weight vector. The three components of this vector correspond to the weights of the coding features of historical load time series data, the coding features of refined grid meteorological historical data, and the coding features of macroeconomic operation data, respectively. The sum of the three components is 1, which ensures the rationality of the fusion process.

[0039] The dynamic fusion unit multiplies the data fusion weight vector element-wise with the corresponding feature channels to obtain the weighted features of each feature channel. Then, the weighted feature channels are concatenated to form the fused feature. For example, the weights of the encoded features of historical load time-series data are... The weights of the coding features of refined grid meteorological historical data are: The weight of the coding features of macroeconomic operation data is The fusion feature is then: .

[0040] After weighted fusion, the fused features are concatenated with the weighted physical simulation feature sequence to form the final input features. The dimension of the final input features is the sum of the dimension of the fused features and the dimension of the physical simulation feature sequence.

[0041] In this embodiment, the core function of the time-series prediction unit is to receive the final input features and output the power load prediction curve for a specified future period through a trained time-series prediction neural network model.

[0042] In this embodiment, the temporal prediction neural network model adopts the temporal fusion Transformer model, which has powerful temporal feature capture and multi-feature fusion capabilities, and can effectively handle long-sequence load prediction problems. The model structure includes an input embedding layer, a multi-head attention mechanism layer, a feedforward neural network layer, and an output layer. The input embedding layer converts the final input features into embedding vectors that the model can process. The multi-head attention mechanism layer captures the temporal dependencies and correlations between features. The feedforward neural network layer performs nonlinear transformations on the features. The output layer outputs the load prediction value for a specified future time period, which can be set to 24 hours, 72 hours, or 168 hours according to actual needs.

[0043] In this embodiment, the hybrid prediction execution module is optimized by combining offline stage joint training with online stage incremental fine-tuning.

[0044] Offline joint training aims to minimize the overall loss of load prediction, simultaneously optimizing the policy network parameters of the dynamic perception control module and the neural network model parameters of the time-series prediction unit through two-layer optimization.

[0045] Inner layer optimization fixed strategy network parameters Optimize the parameters of the time series prediction model The objective function is: ,in To validate the dataset, it consists of historical load data, meteorological data, economic data, power grid parameters, etc., covering different scenario types; To predict the model's loss on the training set, the mean squared error loss function is used to calculate the average of the squared differences between the predicted and actual load values, which measures the model's fit to the training data. The inner optimization uses the Adam optimizer to iteratively update the time-series prediction model parameters. This continues until the loss function converges or the maximum number of iterations is reached. This refers to the "minimum point" operator in mathematics. Indicates network parameters under a fixed policy Given the premise, find the inner optimization objective function. Time series prediction model parameters that reach the minimum value The value of is determined by the core function of the model. Under the current policy network control logic, the optimal parameter configuration of the time series prediction model is determined to ensure that the prediction loss of the model on the training set is minimized, thus providing a basis for subsequent outer layer optimization. yes The result of the calculation represents the optimal time series prediction model parameters obtained from the inner layer optimization. These parameters are based on the fixed policy network parameters. At that time, the prediction loss on the training set is reduced. The minimized parameter combination, including core parameters such as weights and biases of the time series forecasting neural network, directly determines the ability of the time series forecasting model to fit the load change pattern.

[0046] The outer layer optimization is based on the optimal time series prediction model parameters obtained from the inner layer optimization. Optimize network parameters The objective function is: ,in To predict the model's loss on the validation set, the mean squared error loss function is also used to measure the model's generalization ability. The outer layer optimization employs stochastic gradient descent to iteratively update the policy network parameters. This minimizes the model's loss on the validation set, ensuring that the model maintains good predictive performance in different scenarios. This represents the optimal time series prediction model parameters obtained based on inner layer optimization. Find the objective function that enables the outer layer to be optimized. The network parameters of the strategy to reach the minimum value The purpose of this operation is to optimize the control logic of the policy network, enabling the time series prediction model to perform well on the validation set. : yes The result of the calculation represents the optimal policy network parameters obtained from the outer layer optimization. These parameters are based on the optimal time series prediction model parameters of the inner layer. This makes the prediction loss on the validation set... The minimized parameter combination, including core parameters such as the weights and biases of the policy network, is used to generate data fusion control commands and physical guidance control commands during actual system operation. It achieves optimal generalization performance, ensuring that the system can output reliable control commands in different scenarios.

[0047] In this embodiment, the online adaptive fine-tuning module continuously monitors the prediction error output by the hybrid prediction execution module and evaluates the prediction performance in real time.

[0048] The monitoring method involves calculating the absolute or relative error between the predicted load value and the actual load value for each forecast period, and classifying and statistically analyzing the data according to scenario type. Scenario types are categorized based on factors such as weather conditions, date type, and power grid operating status. When the prediction error under a specific scenario type continuously exceeds a set threshold, an online fine-tuning process is triggered. The set threshold is determined based on actual application requirements and is typically between 5% and 10%.

[0049] Online fine-tuning utilizes newly generated data specific to this scenario type to construct a fine-tuning dataset. This dataset includes external environmental data, internal prediction performance metrics, and actual load data for that scenario. The policy network parameters are adjusted through gradient updates in a limited number of steps, typically 10 to 50 steps. This avoids overtraining and model overfitting, while quickly adapting to the characteristics of the specific scenario, thereby improving prediction accuracy within that scenario.

[0050] In this embodiment, the scheduling optimization module receives the power load forecast curve and stability risk index output by the hybrid prediction execution module, and generates generator start-up and shutdown plans and output plans by combining the preset grid operation constraints and power generation cost function.

[0051] Power grid operation constraints include upper and lower limits of generator output, ramp rate constraints, minimum start-up and shutdown time constraints, and power grid transmission capacity constraints. These constraints are set based on power grid equipment parameters and safe operation standards. The power generation cost function includes fuel costs, operation and maintenance costs, and start-up and shutdown costs. It is constructed based on the type and efficiency characteristics of the generator set. For example, the cost function of a coal-fired unit considers the coal consumption rate and the price of coal, while the cost function of a gas-fired unit considers the gas consumption and the price of gas.

[0052] The scheduling optimization module uses a linear programming algorithm to solve for the optimal scheduling scheme. With the goal of minimizing the total power generation cost, it allocates the output of each generating unit under the premise of satisfying the grid operation constraints, determines the start-up and shutdown status of the units and the output value of each time period, and forms a detailed start-up and shutdown plan and output plan to provide guidance for the economic and safe operation of the power system.

[0053] In this embodiment, the system operation flow is as follows: 1. The dynamic perception and control module receives external environmental data and internal predictive performance indicators in real time, preprocesses them to construct a state vector, and inputs it into the policy network to generate a data fusion weight vector and physical guidance strength coefficient; 2. The hybrid prediction execution module receives control commands, and the physical guidance unit determines whether to activate based on the physical guidance intensity coefficient. After activation, it generates a weighted physical simulation feature sequence. 3. The dynamic fusion unit performs weighted fusion of multi-source historical load characteristic data and splices it with the physical simulation characteristic sequence to form the final input features; 4. The time-series forecasting unit processes the final input features through a trained time-series forecasting neural network model and outputs a power load forecast curve and stability risk indicators; 5. The online adaptive fine-tuning module monitors prediction errors and triggers online fine-tuning of the policy network when necessary; 6. The scheduling optimization module generates generator start-up and shutdown plans and output plans based on prediction curves, stability risk indicators, grid operation constraints and generation cost functions.

[0054] This embodiment utilizes a dynamic sensing and control module to perceive changes in the external environment and internal forecast state in real time, dynamically adjusting data fusion weights and physical guidance intensity to enable load forecasting to adapt to the dynamic changes in the value of multi-source data. Simultaneously, it incorporates a physical mechanism simulation model to compensate for the lack of historical data in extreme scenarios, ensuring that the forecast results conform to physical laws. An optimization mechanism combining offline joint training and online incremental fine-tuning enhances the model's generalization and adaptive capabilities. The scheduling optimization module generates scheduling plans based on reliable forecast results, contributing to improved power system operation economy and security, and reducing scheduling decision risks.

[0055] Example 2 like Figures 2 to 3 As shown in the first embodiment, this embodiment elaborates on the specific steps of a machine learning-based power generation load prediction and scheduling system during operation.

[0056] Specifically, the system in this embodiment still includes a dynamic sensing and control module, a hybrid prediction and execution module, an online adaptive fine-tuning module, and a scheduling optimization module. The hardware deployment of each module adopts a distributed architecture, and real-time data interaction is achieved through industrial Ethernet. Data transmission latency is controlled at the millisecond level, meeting the real-time requirements of the power system for load forecasting and scheduling. The following describes the system's workflow in detail, based on actual operating scenarios: Step 1: Data Acquisition and Preprocessing: After the system starts up, it first collects and preprocesses multi-source data to provide basic data support for subsequent prediction and control.

[0057] External environmental data is collected through integration with the existing power grid monitoring system and external public data platforms. Meteorological warning signals are obtained from the regional meteorological department's real-time monitoring system, covering various meteorological disaster warnings such as rainstorms, typhoons, cold waves, and high temperatures. Data is collected every minute to ensure timely capture of dynamic changes in meteorological conditions. Power grid topology parameters and basic load data are obtained from the power grid dispatch automation system. Power grid topology parameters include bus connection relationships, transformer parameters, and line impedance parameters. Basic load data represents the actual electricity load of each region at the current moment, collected every five minutes to ensure data timeliness and accuracy.

[0058] The collection and calculation of internal forecasting performance indicators are based on historical system operating data and real-time forecasting results. Short-term forecasting error statistics are obtained by comparing the load forecast values ​​with the actual load values ​​over several recent forecasting periods. The statistical period is determined according to the forecast duration; hourly forecasts use data from the most recent 24 periods, and daily forecasts use data from the most recent 7 periods. Statistical analysis is used to obtain the distribution characteristics and quantification results of the forecasting error. The deviation between the peak and trough characteristics of the forecast curve and the historical average peak and trough characteristics is calculated by extracting characteristic parameters such as the peak value, trough value, and peak-trough difference of the current forecast curve and comparing them with the average peak and trough characteristic parameters under the same weather conditions and date in the same historical period. Historical data for the same period are selected from the corresponding time periods of the past three years to ensure the rationality of the comparison benchmark.

[0059] All collected data underwent preprocessing. The data cleaning stage removed outliers and missing values. For abnormal fluctuations in meteorological data, interpolation between adjacent time periods was used for correction; for missing values ​​in load data, they were filled in according to load variation patterns within the same time period. The data normalization stage mapped data of different dimensions to the same numerical range, eliminating the impact of dimensional differences on subsequent model calculations. The data encoding stage converted non-numerical data into numerical scalars. For example, meteorological warning signals were converted into corresponding level scalars based on severity, holiday identifiers were converted into corresponding identifier scalars based on date attributes, and the current time was converted into a corresponding time scalar based on time period divisions.

[0060] The second step is the generation of dynamic perception control commands: the preprocessed external environment data and internal predicted performance indicators are input into the dynamic perception control module, which generates data fusion control commands and physical guidance control commands through the policy network.

[0061] The dynamic sensing and control module first constructs a state vector, which consists of an encoded scalar of weather warning level, an uncertainty quantification scalar, a holiday identifier scalar, and a current time scalar. The weather warning level scalar is encoded according to the severity of the weather warning; a lower value is used when there is no warning, and the value gradually increases as the warning level increases, intuitively reflecting the impact of weather conditions on power load. The uncertainty quantification scalar is obtained by calculating the statistical characteristics of prediction errors over the most recent periods, used to quantify the reliability of recent prediction results; a larger value indicates higher prediction uncertainty. The holiday identifier scalar distinguishes between weekdays and statutory holidays; different date types correspond to different fixed values, reflecting the impact of date attributes on load changes. The current time scalar is generated based on the 24-hour time period division, with different values ​​corresponding to different time periods, reflecting the intraday temporal variation pattern of the load.

[0062] The policy network employs a fully connected neural network structure. The input layer receives the constructed state vector, which undergoes nonlinear transformation in the hidden layers before the output layer outputs action commands. These action commands include a data fusion weight vector and a physical guidance strength coefficient. The data fusion weight vector is used to assign weights to each feature channel in the multi-source historical load feature data, while the physical guidance strength coefficient controls the activation state and intensity of the physical guidance unit. The parameters of the policy network have been optimized through offline joint training, enabling it to adaptively generate control commands that conform to the current scenario based on the input state vector, achieving dynamic adaptation to changes in the value of multi-source data and scene characteristics.

[0063] Step 3, Hybrid Prediction Execution: The hybrid prediction execution module receives control commands sent by the dynamic sensing control module, and sequentially completes load prediction through the physical guidance unit, dynamic fusion unit and time-series prediction unit, outputting power load prediction curves and stability risk indicators.

[0064] The physical guidance unit receives the physical guidance intensity coefficient in real time and compares it with a preset activation threshold. When the physical guidance intensity coefficient is greater than the preset activation threshold, the physical guidance unit is activated; when the physical guidance intensity coefficient is not greater than the preset activation threshold, the physical guidance unit remains inactive, and the physical simulation feature sequence is a zero vector and does not participate in subsequent feature concatenation.

[0065] After the physical guidance unit is activated, it invokes a digital twin-based power grid simulation model. This digital twin power grid simulation model is built based on the equipment parameters and operating characteristics of the actual power grid, exhibiting a high degree of consistency with the actual power grid. The model first receives the current power grid topology parameters and basic load data. Then, based on the scenario parameters in the physical guidance control command, it determines the types of extreme scenarios to be simulated, including load surges caused by extreme weather, load transfers caused by power grid faults, and load spikes caused by peak electricity consumption. Subsequently, the model initiates parallel Monte Carlo simulation, using a power system power flow calculation simulation program to solve for the power flow distribution under different extreme scenarios, simulating and generating theoretical load distribution curves corresponding to each scenario. Multiple theoretical load distribution curves are combined to form a physical simulation feature sequence. Before concatenating the physical simulation feature sequence with the output of the dynamic fusion unit, it needs to be multiplied by a physical guidance intensity coefficient for weighting, ensuring that the influence of the physical simulation features matches the requirements of the current scenario, thus strengthening the guiding role of physical mechanisms on the prediction results under extreme scenarios.

[0066] The dynamic fusion unit receives multi-source historical load characteristic data, including historical load time-series data coding features, refined grid meteorological historical data coding features, and macroeconomic operation data coding features. Historical load time-series data coding features are obtained by extracting the trend, periodicity, and randomness characteristics of load data over the past several years, reflecting historical load variation patterns. Refined grid meteorological historical data coding features are obtained by extracting features and spatially pooling data from gridded meteorological data covering the power grid supply area, encompassing the impact of meteorological factors such as temperature, humidity, wind speed, and precipitation on the load. Macroeconomic operation data coding features are obtained by processing economic and social indicators such as regional GDP growth rate, industrial output, the proportion of the tertiary industry, and population, reflecting the long-term impact of macroeconomic development on electricity load.

[0067] The data fusion control command is a data fusion weight vector, where each component corresponds to the weight of each feature channel in the multi-source historical load characteristic data. The dynamic fusion unit multiplies the data fusion weight vector element-wise with the corresponding feature channel to achieve weighted processing of each feature channel. Then, the weighted feature channels are concatenated to form a fused feature. Subsequently, the fused feature is concatenated with the weighted physical simulation feature sequence to obtain the final input feature. This final input feature contains both the statistical regularity of historical data and the constraint information of physical mechanisms, providing comprehensive support for accurate prediction.

[0068] The time-series prediction unit receives the final input features and processes them through a trained time-series prediction neural network model. The time-series prediction neural network model can be either a Long Short-Term Memory (LSTM) network-gated recurrent unit or a time-series fusion Transformer model. This embodiment uses the time-series fusion Transformer model, which possesses powerful time-series feature capture and multi-feature fusion capabilities, effectively handling long-series load forecasting problems. The model obtains optimal parameters through offline joint training optimization. During offline joint training, the strategy network parameters of the dynamic perception control module and the neural network model parameters of the time-series prediction unit are simultaneously optimized to ensure the effectiveness of their collaborative work. After processing the final input features through the model, the time-series prediction unit outputs a power load forecast curve for a specified future time period. The specified time period can be set to 24 hours, 72 hours, or 168 hours according to actual needs. Simultaneously, based on the simulation results of the digital twin power grid simulation model, it outputs a stability risk index for the corresponding scenario's power grid operation. This index reflects the probability of stability problems occurring in the power grid during the forecast period, providing a risk reference for subsequent scheduling optimization.

[0069] Step 4: Online Adaptive Fine-Tuning: The online adaptive fine-tuning module continuously monitors the prediction error output by the hybrid prediction execution module and performs real-time evaluation and dynamic optimization of the system's prediction performance.

[0070] The online adaptive fine-tuning module monitors the deviation between the predicted and actual load values ​​for each forecast period, with the monitoring frequency consistent with the forecast period. During monitoring, forecast errors are categorized and statistically analyzed according to scenario type. Scenario types are defined based on factors such as meteorological conditions, date type, and power grid operating status, such as normal weather weekday scenarios, extreme weather weekday scenarios, and normal weather holiday scenarios. This categorization and statistical analysis allows for the precise identification of forecast performance shortcomings in specific scenarios.

[0071] When the prediction error under a specific scenario type continuously exceeds a set threshold, the system automatically triggers an online fine-tuning process. The set threshold is determined based on the power system's requirements for load prediction accuracy and conforms to industry standards. After the online fine-tuning process is initiated, newly generated operational data for that specific scenario type is first collected, including corresponding external environmental data, internal prediction performance indicators, and actual load data, to construct a dedicated fine-tuning dataset. Then, an incremental fine-tuning method is used to adjust the parameters of the policy network in the dynamic perception control module through a limited number of gradient updates. The number of update steps is strictly controlled during the fine-tuning process to avoid overtraining and model overfitting, while ensuring that the policy network can quickly adapt to the characteristics and patterns of that specific scenario, thereby improving the prediction accuracy under that scenario.

[0072] Step 5: Generation of Dispatch Optimization Plan: The dispatch optimization module receives the power load forecast curve and stability risk index output by the hybrid forecast execution module, and generates generator start-up and shutdown plans and output plans by combining the preset grid operation constraints and generation cost function.

[0073] Power grid operation constraints are set based on power grid equipment parameters and safe operation standards, including upper and lower limits for generator output, ramp rate constraints, minimum start-stop time constraints, and power grid transmission capacity constraints. The upper and lower limits for generator output are determined based on the rated capacity of the units to ensure they operate within a safe output range; ramp rate constraints limit the rate of change in generator output to prevent sudden output fluctuations from affecting power grid stability; minimum start-stop time constraints ensure that the start-stop process meets equipment operating requirements and prevent frequent start-stops from damaging equipment; and power grid transmission capacity constraints are determined based on line transmission capacity to prevent line overload operation.

[0074] The power generation cost function is constructed based on the type and efficiency characteristics of the generator set, encompassing fuel costs, operation and maintenance costs, and start-up and shutdown costs. Fuel costs for coal-fired units are calculated based on coal consumption rate and coal price, while fuel costs for gas-fired units are calculated based on gas consumption and gas price. Operation and maintenance costs are determined based on unit operating time and maintenance standards, and start-up and shutdown costs are determined based on energy consumption and equipment wear during the start-up and shutdown process.

[0075] The scheduling optimization module employs a linear programming algorithm to solve for the optimal scheduling scheme. With the goal of minimizing the total generation cost, and under the premise of satisfying all grid operation constraints, it allocates the output of each generating unit, determining the start-up and shutdown status of each unit and the output value for each time period. The generated generating unit start-up and shutdown plan specifies the start-up and shutdown times of each unit, and the output plan specifies the specific output value of each unit for each time period. The scheduling plan is distributed to each power plant through the grid dispatch management system to guide the actual operation of the generating units, ensuring that the power system achieves economical and safe operation while meeting electricity demand.

[0076] In some alternative implementations, the power system physical mechanism simulation model that runs when the physical guidance unit is activated can be replaced by a power grid simulation model based on electromechanical transients. This model accurately reflects the load change patterns of the power grid under disturbance conditions by simulating the electromechanical transient processes of components such as generators, loads, and transmission lines in the power grid.

[0077] Specifically, this alternative model receives the electromechanical parameters of the current power grid and the scenario parameters from the physical guidance and control commands. The scenario parameters include disturbance information such as line tripping at the location of a short-circuit fault. It then uses an electromechanical transient simulation algorithm to solve for the transient response process of the power grid, generating a theoretical load distribution curve under the disturbance scenario as a physical simulation feature sequence. Simultaneously, the model can also output stability risk indicators such as the power grid transient stability margin, which, along with the power load forecast curve, are provided to the dispatch optimization module.

[0078] The purpose of this implementation is to more accurately simulate the dynamic changes of load in transient disturbance scenarios such as power grid faults, further improve the accuracy of load forecasting under extreme disturbance scenarios, provide a more reliable transient stability risk reference for dispatch optimization, and is suitable for power grid scenarios with high requirements for transient stability.

[0079] In some optional implementations, the multi-source historical load characteristic data received by the dynamic fusion unit can be expanded into four-source characteristic data that includes encoded characteristics of user electricity consumption behavior data. User electricity consumption behavior data is collected from smart meters and covers information such as the user's electricity consumption period, electricity consumption duration, and changes in electricity consumption power, which can reflect the electricity consumption habits of different types of users.

[0080] Specifically, after user electricity consumption behavior data is encoded, it forms independent feature channels. The dimension of the data fusion weight vector generated by the dynamic sensing and control module is expanded accordingly, and the newly added components correspond to the weight allocation of the encoded features of the user electricity consumption behavior data. The dynamic fusion unit multiplies this weight component element-wise with the encoded feature channels of the user electricity consumption behavior data, and then concatenates the weighted results of the other three feature channels to form the fused feature.

[0081] The purpose of this implementation is to introduce the micro-influencing factor of user electricity consumption behavior, which can more accurately capture the individual differences and fine-grained characteristics of load changes. Especially in areas where residential electricity consumption accounts for a high proportion, it can significantly improve the accuracy of load forecasting and make the forecast results more in line with actual electricity demand.

[0082] In some alternative implementations, the time-series prediction neural network model used in the time-series prediction unit can be replaced by a bidirectional gated recurrent unit model. This model, through its gated recurrent structure in both forward and backward directions, can simultaneously capture both the forward and backward time-series dependencies of load data, thus more comprehensively uncovering the time-series patterns of load changes.

[0083] Specifically, the input to this alternative model is the final input features output by the dynamic fusion unit. The feature sequence is processed through a bidirectional gated recurrent structure to extract temporal features in both directions. After mapping by a fully connected layer, the model outputs a power load prediction curve for a specified future period. The model's training also employs a mechanism combining offline joint training with online incremental fine-tuning to ensure coordinated optimization of model parameters and policy network parameters.

[0084] The purpose of this implementation is to make fuller use of the time-series information in the load data with obvious bidirectional time-series correlation, thereby improving the predictive model's ability to capture load change trends. It is suitable for application scenarios with complex load change patterns and strong time-series correlation.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A power generation load forecasting and dispatching system based on machine learning, characterized in that, include: The dynamic perception control module is used to receive external environmental data and internal prediction performance indicators in real time, and generate data fusion control instructions and physical guidance control instructions based on the external environmental data and internal prediction performance indicators. The external environmental data includes at least meteorological warning signals, and the internal prediction performance indicators include at least short-term prediction error statistics. A hybrid prediction execution module, communicatively connected to the dynamic sensing control module, is used to receive the data fusion control command and the physical guidance control command, and to perform load prediction. The hybrid prediction execution module includes: The physical guidance unit is used to activate and run the power system physical mechanism simulation model according to the physical guidance control command when the preset conditions are met, and generate a physical simulation feature sequence. The dynamic fusion unit is used to perform weighted fusion of the input multi-source historical load characteristic data according to the data fusion control command, and to splice the fusion result with the physical simulation characteristic sequence to form the final input feature; The time-series prediction unit is used to receive the final input features and output the power load prediction curve for a specified future period through a trained time-series prediction neural network model.

2. The power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The dynamic perception control module includes a policy network, which uses state vectors. As input, the output includes a data fusion weight vector. With physical guiding strength coefficient The action command, the state vector The construction of the formula is: ,in, This represents a scalar value indicating the coded weather warning level. This represents a scalar quantization of uncertainty calculated based on the prediction errors of the most recent N periods. Indicates the scalar value for holidays. Represents the scalar value at the current moment.

3. A power generation load forecasting and dispatching system based on machine learning according to any one of claims 1 or 2, characterized in that, The physical guiding unit is at the physical guiding strength coefficient Greater than the preset activation threshold It is activated at that time; After activation, the physical guidance unit, based on the current power grid topology parameters and basic load data, calls the power system power flow calculation simulation program to simulate and generate theoretical load distribution curves under at least one extreme scenario, which serve as the physical simulation feature sequence. The physical simulation feature sequence is multiplied by the physical guidance intensity coefficient before being concatenated with the output of the dynamic fusion unit. Weighting is applied.

4. The power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The multi-source historical load characteristic data received by the dynamic fusion unit includes at least the encoding characteristics of historical load time series data, the encoding characteristics of refined grid meteorological historical data, and the encoding characteristics of macroeconomic operation data; The data fusion control instruction is specifically as follows: for the weight allocation vector of the encoded feature channels of different data sources, the dynamic fusion unit realizes the weighted fusion of feature channels by multiplying the weight allocation vector with the corresponding feature channels element by element.

5. A power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The time-series prediction unit uses a time-series prediction neural network model such as a long short-term memory network, a gated recurrent unit, or a time-series fusion Transformer model. The hybrid prediction execution module is optimized by combining offline joint training with online incremental fine-tuning. The offline joint training aims to minimize the overall loss of load prediction and simultaneously optimizes the policy network parameters of the dynamic perception control module and the neural network model parameters of the time series prediction unit.

6. A power generation load forecasting and dispatching system based on machine learning according to claim 5, characterized in that, The offline joint training is achieved through a two-layer optimization, with the outer layer optimizing the objective function. With inner optimization objective function The following relationship must be satisfied: Inner fixed strategy network parameters Optimize the parameters of the time series prediction model for: ； The outer layer optimizes the network parameters based on the results of the inner layer. for: ； in, Represents the validation dataset. To predict the model's loss on the training set, To predict the model's loss on the validation set, This indicates that inner optimization depends on The optimal prediction model parameters.

7. The power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The system also includes an online adaptive fine-tuning module; The online adaptive fine-tuning module continuously monitors the prediction error output by the hybrid prediction execution module. When it detects that the prediction error in a specific scenario type continuously exceeds a set threshold, it triggers online fine-tuning of the policy network in the dynamic perception control module. The online fine-tuning uses newly generated scene data to adjust the policy network parameters through a finite number of gradient updates in order to quickly adapt to the specific scene.

8. The power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The power load forecast curve output by the hybrid forecast execution module is further input into the scheduling optimization module; The scheduling optimization module is used to generate generator start-up and shutdown plans and output plans based on the power load forecast curve and in combination with preset grid operation constraints and generation cost functions.

9. A power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, When generating control commands, the dynamic perception control module also uses internal prediction performance indicators, including the deviation between the peak and valley characteristics of the prediction curve and the average peak and valley characteristics of the same period in history.

10. A power generation load forecasting and dispatching system based on machine learning according to claim 1, characterized in that, The power system physical mechanism simulation model that runs when the physical guidance unit is activated is a power grid simulation model based on digital twins. The digital twin power grid simulation model receives the scenario parameters contained in the physical guidance and control command, performs parallel Monte Carlo simulation, generates the physical simulation feature sequence, and outputs the stability risk index of the power grid operation under the scenario. The stability risk index is output together with the power load forecast curve.

Images