A new energy large model intelligent control and data platform system

Abstract

This invention relates to the field of intelligent operation and control technology for new energy power systems, specifically disclosing a new energy large-scale model intelligent centralized control and data platform system. This invention, through data aggregation and governance, cross-site large-scale model construction, real-time intelligent simulation, and adaptive optimization scheduling modules, performs unified and standardized governance of multi-source heterogeneous new energy power station data. Based on temporal and spatial characteristics, a unified large-scale model is trained and a parallel inference interface is provided. Integrated analysis, simulation, and dynamic optimization scheduling of operational data are then performed. This solves the problems of data silos, reliance on manual experience leading to response lags and insufficient prediction accuracy, and a lack of integrated large-scale model capabilities, thereby improving the collaborative optimization and scheduling response capabilities of new energy power stations.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and control technology for new energy power systems, and in particular to an intelligent centralized control and data platform system for a large-scale new energy model. Background Technology

[0002] With the rapid development of the new energy industry and the large-scale grid connection of distributed energy sources such as wind power and photovoltaics, higher demands are placed on the intelligence level of centralized control systems. Existing new energy centralized control platforms mostly adopt traditional centralized or decentralized architectures, with data from each site stored independently and using inconsistent standards, forming data silos and making it difficult to achieve cross-domain collaborative analysis and global optimization scheduling. At the same time, traditional centralized control systems rely on human experience and fixed rules for monitoring and decision-making. Faced with massive amounts of time-series data and complex operating conditions, they suffer from problems such as response lag and insufficient prediction accuracy, failing to effectively cope with the randomness and volatility of new energy output. Although big data and artificial intelligence technologies are gradually being applied, there is a lack of integrated platforms for large-scale model training and inference for new energy scenarios. Data preprocessing, feature engineering, and model deployment processes are fragmented, data value mining is insufficient, the level of system intelligence is limited, and model update and iteration cycles are long, making it difficult to support the efficient absorption and stable operation of new energy.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent centralized control and data platform system for a large-scale new energy model. The technical solution of this system is as follows:

[0005] The data aggregation and governance module is used to collect various types of heterogeneous data from multiple new energy power stations, aggregate, clean and standardize the various types of heterogeneous data, and output standardized time-series data with unified specifications.

[0006] The cross-site large model construction module is used to construct and train a unified large model in the new energy field based on the time series features and site spatial correlation features in the standardized time series data, and output the trained large model ontology and multiple parallel task inference interfaces.

[0007] The real-time intelligent simulation module is used to receive the current station operation data and call the large model ontology and the multiple parallel task inference interfaces to perform integrated analysis and simulation of the current station operation data, and generate real-time simulation results including ultra-short-term power prediction, equipment status diagnosis and risk warning signals.

[0008] The adaptive optimization scheduling module is used to receive the real-time simulation results and external power grid scheduling instructions, perform dynamic optimization calculations based on the real-time simulation results and external power grid scheduling instructions, and generate and issue an optimized scheduling instruction set to coordinate the operation of the multiple new energy power stations.

[0009] Furthermore, the data aggregation and governance module is also used to: perform real-time quality scoring on the standardized time-series data that has completed standardized governance, and the quality score is used for data weighting and confidence identification in subsequent processing.

[0010] Furthermore, the cross-site large model construction module is also used to: employ a meta-learning framework to pre-train the new energy field large model using multiple historical site datasets to generate initial parameters with rapid cross-scenario adaptability; the new energy field large model is constructed based on a spatiotemporal graph neural network architecture.

[0011] Furthermore, in the spatiotemporal graph neural network architecture, the dynamic association weights between new energy power stations are generated by a learnable association network module, and the input of the association network module is the real-time feature sequence and historical interaction features of the corresponding new energy power station nodes.

[0012] Furthermore, the real-time intelligent inference module is also used to: perform real-time calibration and data assimilation of the internal hidden state of the large model ontology based on the residual between the latest measured data and the real-time inference results.

[0013] Furthermore, the dynamic optimization calculation performed by the adaptive optimization scheduling module also quantifies the risk warning signal into uncertainty parameters related to the scheduling safety boundary, and introduces them into the optimization model as opportunity constraints.

[0014] Furthermore, the adaptive optimization scheduling module adopts a two-layer rolling time-domain optimization framework. The upper-layer model solves for the power reference values ​​of multiple power stations, and the lower-layer model generates the optimized scheduling instruction set based on the power reference values, the uncertainty parameters, and the internal constraints of the power stations.

[0015] Furthermore, when the cross-site large model construction module trains the large model in the new energy field, the loss function used is a composite loss function that integrates data quality weights, spatiotemporal consistency regularization terms, and prediction uncertainty measures.

[0016] The expression for the composite loss function is:

[0017]

[0018] in, This represents the total loss value. This represents the total number of training samples. Indicates the first The data quality score weights corresponding to each sample Indicates the first The true value of each sample Indicates the first The model prediction value for each sample. The model represents the first The variance of the uncertainty in the prediction of a single sample. For uncertainty regularization coefficients, Represents the trace operation of a matrix. This is the feature representation matrix of all new energy power station nodes in the model's hidden layer. The normalized graph Laplacian matrix is ​​constructed based on the dynamic association weights. is the spatiotemporal consistency regularization coefficient.

[0019] Furthermore, the optimization objective of the upper-level model is a robust objective function that considers expected returns and risk costs in the worst case.

[0020] The expression for the robustness objective function is:

[0021]

[0022] in, This indicates the target value to be optimized. This represents the planned output vector. This represents the feasible region defined by fundamental physical constraints. Represents uncertainty variables The mathematical expectation, This characterizes the prediction uncertainty and equipment state randomness inherent in the real-time simulation results. for The confidence set, To consider the revenue function of market electricity prices and generation costs, This is the risk aversion coefficient. This indicates the total number of time periods to be optimized. Indicates the number of key safety indicators. Indicates the first Time period The deviation of a safety indicator from its dynamic safety boundary. Let be a non-negative convex penalty function with respect to the deviation amount.

[0023] Furthermore, the system also includes:

[0024] The online model evolution module is used to dynamically sparsify and add / delete structured network connections of the large model ontology based on the state correction amount and long-term performance evaluation generated by the real-time calibration and data assimilation process.

[0025] The technical solution of this invention performs unified and standardized governance of multi-source heterogeneous new energy power station data through data aggregation and governance, cross-station large model construction, real-time intelligent inference and adaptive optimization scheduling modules. It trains a unified large model based on temporal and spatial characteristics and provides a parallel inference interface to perform integrated analysis, inference and dynamic optimization scheduling of operation data. This solves the problems of data silos, reliance on human experience leading to response lag and insufficient prediction accuracy, and lack of integrated large model capabilities, and realizes the improvement of collaborative optimization and scheduling response capabilities of new energy power stations.

[0026] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0027] 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.

[0028] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0029] Figure 1 This is a schematic diagram of an embodiment of a new energy large-scale intelligent centralized control and data platform system according to the present invention. Detailed Implementation

[0030] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0031] Figure 1 This diagram illustrates the structure of an embodiment of a new energy large-scale intelligent centralized control and data platform system provided by the present invention. Figure 1 As shown, the system includes:

[0032] The data aggregation and governance module 110 is used to collect various types of heterogeneous data from multiple new energy power stations, aggregate, clean and standardize the various types of heterogeneous data, and output standardized time-series data with unified specifications.

[0033] Among them, "new energy power stations" refers to power plant clusters that generate electricity using renewable energy sources such as wind and solar power; for example, wind farm A and photovoltaic power station B managed by a centralized control center. "Multiple types of heterogeneous data" refers to data sets with different structures, formats, and semantics generated during the new energy power generation process; for example, SCADA power data from wind farm A, inverter status alarm logs from photovoltaic power station B, and numerical weather forecast data from public services. "Standardized time-series data" refers to time-series data with a unified time base and standardized numerical format formed after cleaning, aligning, and transforming the original heterogeneous data; for example, unifying the second-level wind speed data from wind farm A and the minute-level irradiance data from photovoltaic power station B into a data sequence with timestamp alignment, a sampling interval of 15 minutes, and normalized values.

[0034] The cross-site large model construction module 120 is used to construct and train a unified large model in the new energy field based on the time series features and site spatial correlation features in the standardized time series data, and output the trained large model ontology and multiple parallel task inference interfaces.

[0035] Among these, temporal features refer to patterns or statistics extracted from time-series data that characterize the evolution of data over time; for example, the trend component, periodic component, and fluctuation amplitude extracted from the power sequence of wind farm A over the past 24 hours. Spatial correlation features of wind farms refer to characteristics representing the mutual influence and dependence between different renewable energy power plants due to geographical, climatic, or grid connection relationships; for example, wind farms A and C, located in the same wind belt, have high synchronous correlation in their power output sequences. Large-scale models in the renewable energy field refer to complex parametric models trained on massive amounts of renewable energy operation data using deep learning technology, capable of performing various analytical tasks in the renewable energy field; for example, a neural network model based on the Transformer architecture trained on historical data from numerous wind farms and photovoltaic power plants, with a parameter scale reaching billions. Large model ontology refers to the model entity containing all executable parameters and computational graph structure after training; for example, neural network weights and architecture definitions saved as a specific format file (such as .pt or .pb). Multiple parallel task inference interfaces refer to application programming interfaces that are developed based on the same large model ontology and can handle different analysis tasks simultaneously; for example, an API that provides three functions: power prediction, device health assessment, and anomaly detection, can receive requests and return results at the same time.

[0036] The real-time intelligent simulation module 130 is used to receive the current station operation data and call the large model ontology and the multiple parallel task inference interfaces to perform integrated analysis and simulation of the current station operation data, and generate real-time simulation results including ultra-short-term power prediction, equipment status diagnosis and risk warning signals.

[0037] The current station operation data refers to the real-time operation data actually measured from the renewable energy power station within the most recent data acquisition cycle; for example, the wind turbine speed of wind farm A and the DC-side voltage of photovoltaic power station B collected at time t. Integrated analysis and extrapolation refers to the process of using a large model to perform a forward calculation on the input data and simultaneously outputting multiple related analysis results; for example, inputting the station operation data into a large model, which simultaneously outputs future power curves, equipment health scores, and system risk indicators. Ultra-short-term power prediction refers to estimating the power generation of renewable energy power stations for the next few minutes to hours; for example, predicting the output value of wind farm A in the next 15 minutes, 30 minutes, 45 minutes, and 60 minutes. Equipment status diagnosis refers to the process of determining the current health status and operating mode of renewable energy power generation equipment based on operational data analysis; for example, analyzing the electrical parameters of the inverter in photovoltaic power station B to determine whether it has insulation aging and whether it is currently operating under power limiting conditions. Risk warning signals refer to event identifiers identified by the system that characterize potential safety risks or operational anomalies; for example, a "gearbox vibration exceeding limit warning" or a "collection line power prediction exceeding limit warning" generated by the system. Real-time simulation results refer to the comprehensive output data set generated during the integrated analysis and simulation process; for example, a data package containing future power prediction values, equipment status score lists, and risk warning signal lists.

[0038] The adaptive optimization scheduling module 140 is used to receive the real-time simulation results and external power grid scheduling instructions, perform dynamic optimization calculations based on the real-time simulation results and external power grid scheduling instructions, and generate and issue an optimized scheduling instruction set to coordinate the operation of the multiple new energy power stations.

[0039] External grid dispatch instructions refer to control commands from the grid dispatching agency that require adjustments to the total output of renewable energy sources; for example, a command requiring the total output of the central control center to be controlled between 100MW and 120MW within the next hour. Optimized dispatch instruction sets refer to a collection of specific control commands generated to coordinate the operation of multiple renewable energy plants to achieve specific objectives; for example, a command to reduce output sent to wind farm A, or a command to adjust the power factor sent to photovoltaic power plant B.

[0040] The technical solution in this embodiment uses data aggregation and governance, cross-site large model construction, real-time intelligent inference and adaptive optimization scheduling modules to perform unified and standardized governance of multi-source heterogeneous new energy power station data. It trains a unified large model based on temporal and spatial characteristics and provides a parallel inference interface to perform integrated analysis, inference and dynamic optimization scheduling of operational data. This solves the problems of data silos, reliance on human experience leading to response lag and insufficient prediction accuracy, and lack of integrated large model capabilities, and realizes the improvement of collaborative optimization and scheduling response capabilities of new energy power stations.

[0041] In an optional manner, the data aggregation and governance module 110 is further configured to: perform real-time quality scoring on the standardized time-series data that has undergone standardized governance, wherein the quality score is used for data weighting and confidence identification in subsequent processing.

[0042] The quality score refers to a numerical indicator obtained by quantitatively evaluating the accuracy, completeness, and timeliness of the data; for example, a score of 0.92 (out of 1.0) is obtained after evaluating a batch of power data.

[0043] Among the above-mentioned optional methods, the standardized time series data are further subjected to real-time quality scoring to provide data weighting basis and confidence indicators for subsequent processing, thereby improving data reliability and model training effect.

[0044] In an optional manner, the cross-site large model construction module 120 is further configured to: employ a meta-learning framework to pre-train the new energy field large model using multiple historical site datasets to generate initial parameters with rapid cross-scenario adaptability; the new energy field large model is constructed based on a spatiotemporal graph neural network architecture.

[0045] In this context, a meta-learning framework refers to a machine learning approach that enables models to quickly adapt to new tasks by leveraging prior task experience. For example, the MAML algorithm allows models to be pre-trained on multiple wind and solar power plant datasets, gaining the ability to rapidly adapt to new plants. Historical plant datasets refer to collections of data consisting of historical operational data from renewable energy plants, used for model training or analysis. For example, the average hourly power, wind speed, and temperature data recorded by wind farm A over the past three years. Initial parameters refer to the initial values ​​of model parameters that possess cross-task generalization capabilities before fine-tuning the model for a specific task; for example, model weights obtained after meta-learning pre-training.

[0046] Among the above-mentioned optional methods, a meta-learning framework is further adopted for pre-training, which enables large models in the new energy field to have the ability to adapt quickly across scenarios, shorten the deployment cycle of new sites and improve the generalization performance of the model.

[0047] In one alternative approach, in the spatiotemporal graph neural network architecture, the dynamic association weights between new energy power stations are generated by a learnable association network module, the input of which is the real-time feature sequence and historical interaction features of the corresponding new energy power station nodes.

[0048] The "association network module" refers to a sub-neural network component within the larger model specifically designed for dynamically calculating the correlation weights between power plants; for example, a small neural network that takes the features of two power plants as input and outputs a scalar value representing the strength of their correlation. The "real-time feature sequence" refers to a data sequence generated by new energy power plants within a recent continuous time period, reflecting their transient operating status; for example, the generator winding temperature sequence collected per second for wind farm A over the past 10 minutes. The "historical interaction features" refer to statistical features extracted from long-term historical data between power plants, characterizing their mutual influence patterns; for example, the calculated Pearson correlation coefficient between the daily output curves of wind farm A and photovoltaic power plant B.

[0049] In the above-mentioned optional methods, dynamic correlation weights between new energy power stations can be generated by a learnable correlation network module, and the spatial correlation feature capture capability can be improved by using real-time feature sequences and historical interaction features for modeling.

[0050] In an alternative approach, the real-time intelligent inference module 130 is further configured to: perform real-time calibration and data assimilation of the internal hidden state of the large model ontology based on the residual between the latest measured data and the real-time inference result.

[0051] The latest measured data refers to the raw measurement value directly acquired and uploaded by the field sensing device at the most recent data acquisition time; for example, the current value collected from the B combiner box of the photovoltaic power station at time t. The internal hidden state refers to the internal variables used by the neural network model to remember historical information during the processing of sequential data; for example, the hidden layer vector calculated by a recurrent neural network unit at time t.

[0052] Among the above-mentioned optional methods, the internal hidden states of the large model ontology are further calibrated and assimilated in real time based on the residuals of the latest measured data and real-time inference results, thereby improving prediction accuracy and model robustness.

[0053] In an alternative approach, the dynamic optimization calculation performed by the adaptive optimization scheduling module 140 also quantifies the risk warning signal into uncertainty parameters related to the scheduling safety boundary, and introduces them as opportunity constraints into the optimization model.

[0054] Uncertainty parameters refer to numerical variables used to quantify uncertainties in model predictions or system decisions; for example, the standard deviation σ accompanying power prediction results, or the adjustment δ used in safety boundary calculations. Chance constraints refer to mathematical expressions in optimization problems that allow constraints to be satisfied with a certain probability; for example, requiring a probability of the system's total active power output being greater than 100MW of no less than 95%.

[0055] In the above-mentioned optional methods, the risk warning signal is further quantified into uncertainty parameters related to the scheduling safety boundary, and introduced into the optimization model as an opportunity constraint to improve the safety and reliability of the scheduling scheme.

[0056] In one alternative approach, the adaptive optimization scheduling module 140 employs a two-layer rolling time-domain optimization framework. The upper-layer model solves for the power reference values ​​of multiple power stations, while the lower-layer model generates the optimized scheduling instruction set based on the power reference values, the uncertainty parameters, and the internal constraints of the power stations.

[0057] The dual-layer rolling time-domain optimization framework refers to a decision-making framework that performs optimization calculations in two layers, rolling over time. For example, the upper layer optimizes the power baseline for the next 4 hours every 15 minutes, while the lower layer optimizes detailed control instructions for the next 15 minutes based on the upper layer's results every 1 minute. The upper-layer model refers to the optimization model responsible for macro-level resource allocation on a longer time scale within the dual-layer optimization framework; for example, a model optimizing the daily power generation plan for each station with an hourly time granularity. The multi-station power baseline value refers to the planned output reference value allocated to each new energy station, determined by the upper-layer model; for example, the planned output value of 80MW allocated to wind farm A in the next time period. The lower-layer model refers to the optimization model responsible for refined control on a short time scale within the dual-layer optimization framework; for example, a model optimizing the specific pitch angle instructions for each wind turbine with a minute-level time granularity. Internal constraints of a station refer to the physical or technical limitations that a single new energy station or equipment must follow for operation; for example, the minimum cut-in wind speed and maximum allowable rotational speed of a single wind turbine.

[0058] Among the above-mentioned optional methods, a two-layer rolling time-domain optimization framework is further adopted. The upper-layer model solves for the power reference values ​​of multiple power stations, and the lower-layer model generates an optimized scheduling instruction set by combining the internal constraints of the power stations, thereby improving the scheduling calculation efficiency and instruction executability.

[0059] In one optional approach, when the cross-site large model construction module trains the large model in the new energy field, the loss function used is a composite loss function that integrates data quality weights, spatiotemporal consistency regularization terms, and prediction uncertainty measures.

[0060] The expression for the composite loss function is:

[0061]

[0062] in, This represents the total loss value. This represents the total number of training samples. Indicates the first The data quality score weights corresponding to each sample Indicates the first The true value of each sample Indicates the first The model prediction value for each sample. The model represents the first The variance of the uncertainty in the prediction of a single sample. For uncertainty regularization coefficients, Represents the trace operation of a matrix. This is the feature representation matrix of all new energy power station nodes in the model's hidden layer. The normalized graph Laplacian matrix is ​​constructed based on the dynamic association weights. is the spatiotemporal consistency regularization coefficient.

[0063] It should be noted that the construction principle of the composite loss function lies in solving the problem of multi-objective collaborative optimization in the training of large models in the new energy field. This function is not simply a superposition of mean squared errors, but rather a structural integration of three core considerations. The first term, weighted mean squared error, ensures that the model's predicted values ​​are close to the true values ​​and introduces data quality weights, giving higher weight to high-quality data during training. The second term incorporates the idea of ​​uncertainty quantification, including a log-variance term and a standardized error term. The former encourages the model to output a larger variance for uncertain samples, while the latter constrains the variance estimate from increasing indefinitely. Together, they enable the model to learn and output a meaningful measure of predictive uncertainty. The third term, spatiotemporal consistency regularization, is based on graph signal processing theory. It applies a smoothing constraint to the hidden feature representations of all station nodes in the model through the graph Laplacian matrix, forcing geographically or electrically adjacent station nodes to be close to each other in the feature space, thereby ensuring that the model's inference results conform to physical laws in spatial distribution. Functionally, this composite loss function simultaneously guides the model to optimize prediction accuracy, calibrate prediction uncertainty, and maintain the spatiotemporal rationality of cross-site extrapolation results. It is the core training mechanism for achieving high-precision and interpretable cross-site integrated analysis.

[0064] Among the above-mentioned optional methods, a composite loss function that integrates data quality weights, spatiotemporal consistency regularization terms, and prediction uncertainty measures is further adopted for model training to improve training performance and spatiotemporal consistency constraint strength.

[0065] In one alternative approach, the optimization objective of the upper-level model is a robust objective function that considers expected returns and worst-case risk costs;

[0066] The expression for the robustness objective function is:

[0067]

[0068] in, This indicates the target value to be optimized. This represents the planned output vector. This represents the feasible region defined by fundamental physical constraints. Represents uncertainty variables The mathematical expectation, This characterizes the prediction uncertainty and equipment state randomness inherent in the real-time simulation results. for The confidence set, To consider the revenue function of market electricity prices and generation costs, This is the risk aversion coefficient. This indicates the total number of time periods to be optimized. Indicates the number of key safety indicators. Indicates the first Time period The deviation of a safety indicator from its dynamic safety boundary. Let be a non-negative convex penalty function with respect to the deviation amount.

[0069] It should be noted that the robust objective function is constructed to achieve an optimal balance between economic efficiency and safety in scheduling decisions, addressing the dual uncertainties of new energy output and market environment. This function employs the mathematical form of a partially robust optimization framework. Its core structure consists of two parts. The first is the expected return function, which, under uncertain variables... The first term is the average economic benefit that the planned output vector P can bring within a certain confidence set, reflecting the decision-maker's pursuit of basic economic efficiency. The second term is the risk cost term, which calculates the sum of penalties for deviations of the system safety indicators from their dynamic boundaries under the worst-case scenario that may occur within the same uncertainty confidence set. This penalty is mapped to the deviation of each safety indicator through a non-negative convex function. The two terms are connected by a risk aversion coefficient. A trade-off is made. Functionally, the objective function drives the optimization model not only to seek high average returns but also to proactively guard against potential safety risks under extremely unfavorable scenarios. It transforms the predictive uncertainties and risk warning signals inherent in real-time simulation results into a set of uncertainty parameters in the optimization problem. The dynamic safety boundary enables the generated scheduling instruction set to have inherent robustness, ensuring the stable operation of the power grid in complex and ever-changing environments.

[0070] Among the above-mentioned alternative approaches, a robust objective function that considers expected returns and worst-case risk costs is further adopted as the optimization objective of the upper-level model to improve the risk resistance and economic balance of the scheduling strategy.

[0071] In an alternative embodiment, the system further includes:

[0072] The online model evolution module is used to dynamically sparsify and add / delete structured network connections of the large model ontology based on the state correction amount and long-term performance evaluation generated by the real-time calibration and data assimilation process.

[0073] State correction refers to the adjustment value calculated to reduce the deviation between the model's internal state and the actual system state; for example, the correction vector for the model's hidden state calculated by Kalman filtering based on the difference between the latest measured power and the model's projected power. Long-term performance evaluation refers to the systematic measurement and evaluation of the model's overall performance over an extended time period; for example, statistical indicators such as the model's average prediction error and early warning accuracy over the past 30 days. Network connection refers to the directed links between different nodes (neurons) in an artificial neural network, typically with trainable weight parameters; for example, the weight matrix of fully connected nodes between adjacent layers in a feedforward neural network. Dynamic sparsity refers to dynamically removing less important connections in a neural network based on online evaluation results; for example, setting connections whose absolute weight value is consistently below a threshold ε to zero. Structured addition and deletion refers to systematically adding or removing groups of nodes or connection patterns in a neural network based on performance evaluation and knowledge requirements; for example, adding a dedicated feature extraction submodule to the model to improve the ability to identify a new fault mode.

[0074] In the above-mentioned optional methods, the model online evolution module is further used to dynamically sparsify and add or delete network connections of the large model ontology based on state correction amount and long-term performance evaluation, so as to achieve adaptive optimization and lightweighting of the model.

[0075] In another embodiment of the present invention, it specifically includes:

[0076] S10: The data aggregation and governance module collects various types of heterogeneous data from multiple new energy power plants, including power data, equipment status logs and numerical weather forecast data. It aggregates, cleans and corrects outliers of various types of heterogeneous data, performs timestamp alignment and resampling operations, generates standardized time series data with a unified time base and standardized numerical format, and performs real-time quality scoring on the standardized time series data.

[0077] S20: The cross-site large model construction module receives standardized time-series data, uses a meta-learning framework to pre-train a large model in the new energy field using multiple historical site datasets, generates initial parameters with rapid cross-scenario adaptability, defines a spatiotemporal graph neural network architecture based on time-series features and site spatial correlation features, wherein the dynamic correlation weights between new energy sites are generated by the learnable correlation network module based on real-time feature sequences and historical interaction features, and uses a composite loss function that integrates data quality weights, spatiotemporal consistency regularization terms and prediction uncertainty measures to train the model, and outputs the trained large model ontology and multiple parallel task inference interfaces;

[0078] S30: The real-time intelligent simulation module receives the current station operation data, calls the large model ontology and multiple parallel task inference interfaces to perform integrated analysis and simulation, and synchronously generates real-time simulation results including ultra-short-term power prediction, equipment status diagnosis and risk warning signals. Based on the residual between the latest measured data and the real-time simulation results, it performs real-time calibration and data assimilation of the hidden state inside the large model ontology.

[0079] S40: The adaptive optimization scheduling module receives real-time simulation results and external power grid scheduling instructions, quantifies risk warning signals into uncertainty parameters related to the scheduling safety boundary and uses them as opportunity constraints. It adopts a two-layer rolling time-domain optimization framework. The upper-layer model solves for the power benchmark values ​​of multiple power stations using a robust objective function that considers expected returns and risk costs in the worst case. The lower-layer model performs rolling optimization based on the power benchmark values, uncertainty parameters and internal constraints of the power stations, and generates and issues an optimized scheduling instruction set that coordinates the operation of multiple new energy power stations.

[0080] S50: The online model evolution module performs dynamic sparsification to remove low-weight connections in the network connections of the large model ontology based on the state correction amount generated during real-time calibration and data assimilation processes and the long-term performance evaluation of model prediction error and early warning accuracy indicators. It also performs structured additions and deletions based on newly emerging operating modes and knowledge requirements to optimize the network architecture and achieve adaptive continuous optimization of the model.

[0081] This embodiment addresses the core technical challenges of existing technologies by implementing data aggregation and standardized governance, constructing a large cross-site model based on meta-learning and spatiotemporal graph neural networks, combining real-time state calibration with integrated intelligent extrapolation, incorporating risk quantification and two-layer robust optimization for adaptive scheduling, and online model evolution based on long-term evaluation. These challenges include data silos, weak model generalization ability, response lag and insufficient prediction accuracy due to reliance on human experience, poor risk response capability of scheduling strategies, and long model update cycles. The embodiment achieves deep integration and high-quality utilization of new energy power station data value, significant improvement in cross-scenario collaborative analysis and extrapolation accuracy, balanced optimization of scheduling decision-making economy and safety reliability, and adaptive and continuous enhancement of the system's overall intelligent level.

[0082] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0083] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and do not imply a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0084] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A smart centralized control and data platform system for a large-scale new energy model, characterized in that, The system includes: The data aggregation and governance module is used to collect various types of heterogeneous data from multiple new energy power stations, aggregate, clean and standardize the various types of heterogeneous data, and output standardized time-series data with unified specifications. The cross-site large model construction module is used to construct and train a unified large model in the new energy field based on the time series features and site spatial correlation features in the standardized time series data, and output the trained large model ontology and multiple parallel task inference interfaces. The real-time intelligent simulation module is used to receive the current station operation data and call the large model ontology and the multiple parallel task inference interfaces to perform integrated analysis and simulation of the current station operation data, and generate real-time simulation results including ultra-short-term power prediction, equipment status diagnosis and risk warning signals. An adaptive optimization scheduling module is used to receive the real-time simulation results and external power grid scheduling instructions, perform dynamic optimization calculations based on the real-time simulation results and external power grid scheduling instructions, and generate and issue an optimized scheduling instruction set to coordinate the operation of the multiple new energy power stations. The cross-site large model construction module is also used to: employ a meta-learning framework to pre-train the new energy field large model using multiple historical site datasets to generate initial parameters with rapid cross-scenario adaptability; the new energy field large model is constructed based on a spatiotemporal graph neural network architecture; In the spatiotemporal graph neural network architecture, the dynamic association weights between new energy power stations are generated by a learnable association network module. The input of the association network module is the real-time feature sequence and historical interaction features of the corresponding new energy power station nodes. When the cross-site large model construction module trains the large model in the new energy field, the loss function used is a composite loss function that integrates data quality weights, spatiotemporal consistency regularization terms, and prediction uncertainty measures. The expression for the composite loss function is: in, This represents the total loss value. This represents the total number of training samples. Indicates the first The data quality score weights corresponding to each sample Indicates the first The true value of each sample Indicates the first The model prediction value for each sample. The model represents the first The variance of the uncertainty in the prediction of a single sample. For uncertainty regularization coefficients, Represents the trace operation of a matrix. This is the feature representation matrix of all new energy power station nodes in the model's hidden layer. The normalized graph Laplacian matrix is ​​constructed based on the dynamic association weights. is the spatiotemporal consistency regularization coefficient.

2. The intelligent centralized control and data platform system for large-scale new energy models according to claim 1, characterized in that, The data aggregation and governance module is also used to: perform real-time quality scoring on the standardized time-series data that has completed standardized governance, and the quality score is used for data weighting and confidence identification in subsequent processing.

3. The intelligent centralized control and data platform system for large-scale new energy models according to claim 1, characterized in that, The real-time intelligent inference module is also used to: perform real-time calibration and data assimilation of the internal hidden state of the large model ontology based on the residual between the latest measured data and the real-time inference results.

4. The intelligent centralized control and data platform system for large-scale new energy models according to claim 1, characterized in that, The adaptive optimization scheduling module performs dynamic optimization calculations and quantifies the risk warning signal into uncertainty parameters related to the scheduling safety boundary, which are then introduced into the optimization model as opportunity constraints.

5. The intelligent centralized control and data platform system for large-scale new energy models according to claim 4, characterized in that, The adaptive optimization scheduling module adopts a two-layer rolling time-domain optimization framework. The upper-layer model solves for the power reference values ​​of multiple power stations, and the lower-layer model generates the optimized scheduling instruction set based on the power reference values, the uncertainty parameters, and the internal constraints of the power stations.

6. The intelligent centralized control and data platform system for large-scale new energy models according to claim 5, characterized in that, The optimization objective of the upper-level model is a robust objective function that considers expected returns and risk costs in the worst case. The expression for the robustness objective function is: in, This indicates the target value to be optimized. This represents the planned output vector. This represents the feasible region defined by fundamental physical constraints. Represents uncertainty variables The mathematical expectation, This characterizes the prediction uncertainty and equipment state randomness inherent in the real-time simulation results. for The confidence set, To consider the revenue function of market electricity prices and generation costs, This is the risk aversion coefficient. This indicates the total number of time periods to be optimized. Indicates the number of key safety indicators. Indicates the first Time period The deviation of a safety indicator from its dynamic safety boundary. Let be a non-negative convex penalty function with respect to the deviation amount.

7. The intelligent centralized control and data platform system for large-scale new energy models according to claim 3, characterized in that, The system also includes: The online model evolution module is used to dynamically sparsify and add / delete structured network connections of the large model ontology based on the state correction amount and long-term performance evaluation generated by the real-time calibration and data assimilation process.

Images