New energy centralized control center power prediction and optimization decision system

Abstract

The invention relates to the technical field of new energy power generation management, and specifically provides a new energy centralized control center power prediction and optimization decision system. The system comprises a data acquisition and preprocessing module, a core power prediction module and an optimization decision execution module. The core power prediction module adopts a bidirectional time-space neural network model based on an attention mechanism, extracts time sequence features of each station through a bidirectional time sequence coding layer in sequence, calculates dynamic weights through a graph attention network in a spatial feature aggregation layer so as to aggregate spatial features of adjacent stations, and finally fuses meteorological data to output a power prediction value. And the optimization decision execution module constructs an optimization model considering the risk cost by using the prediction result and the uncertainty information of the prediction result so as to generate a power control instruction. According to the method, the precision of centralized control center-level power prediction and the robustness of optimization decision are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation management technology, and more specifically, to a power prediction and optimization decision-making system for a new energy centralized control center. Background Technology

[0002] With the rapid increase in the proportion of intermittent renewable energy sources such as wind and solar power in the power system, the renewable energy control center, as the core of multiple distributed power plants within its management area, plays a crucial role in the safe and stable operation of the power grid through accurate power forecasting and reliable optimization decisions. Accurate power forecasting is fundamental to formulating scientific power generation plans, alleviating grid peak-shaving pressure, and promoting the efficient consumption of renewable energy. Therefore, developing advanced power forecasting and optimization decision-making systems has significant practical implications.

[0003] Existing renewable energy power prediction technologies mostly focus on independent predictions for individual power plants or use simple arithmetic summation methods for centralized control center-level predictions, failing to fully consider the spatiotemporal correlations between power plants. Regarding prediction models, traditional time-series models or shallow machine learning models have limited ability to characterize complex nonlinear relationships, while some early deep learning models, although capable of extracting temporal features, struggle to effectively integrate spatial information. Specifically, existing technologies typically treat each power plant as an independent entity, ignoring the power linkage effects between geographically proximate power plants due to similar meteorological conditions. In data processing, the utilization of multi-source information such as numerical weather forecasts is relatively crude, lacking deep feature alignment and filtering mechanisms. Furthermore, most systems treat prediction and optimization as two independent processes; the uncertainty information output by the prediction module is not effectively utilized by the optimization decision-making module, resulting in insufficient risk resilience of the decision-making scheme when wind and solar power fluctuations increase, making it difficult to meet the increasingly stringent performance requirements of the power grid.

[0004] Therefore, a power prediction and optimization decision-making system for a new energy centralized control center is proposed to address the above problems. The core issues to be solved are: how to overcome the limitations of existing technologies in deeply mining the spatiotemporal correlation characteristics of new energy power plant clusters; how to improve the efficiency of prediction models in integrating and utilizing multi-source heterogeneous data; and how to effectively transmit the uncertainty risks identified in the prediction stage to the optimization decision-making stage to generate more robust power control commands. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a power prediction and optimization decision system for a new energy centralized control center to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a power prediction and optimization decision-making system for a new energy centralized control center, comprising: The data acquisition and preprocessing module is used to acquire historical power data, real-time operation data and numerical weather forecast data of each new energy power station under the central control center, and to perform quality verification and normalization processing on the data to form a standardized training sample set. The core power prediction module, connected to the data acquisition and preprocessing module, integrates a bidirectional spatiotemporal neural network model based on an attention mechanism. This model processes the training sample set and outputs power prediction values ​​for a specific future time period. The bidirectional spatiotemporal neural network model sequentially includes: The bidirectional temporal coding layer employs a bidirectional gated recurrent unit network to extract the temporal features of the power sequence of each new energy power station in both forward and reverse order, and outputs a hidden state sequence containing contextual information. The spatial feature aggregation layer, connected to the bidirectional temporal coding layer, adopts a graph attention network. By calculating the dynamic attention weights between any two stations, it aggregates the spatial features of adjacent stations and generates an enhanced spatial feature representation for each station. The multi-source fusion prediction layer, connected to the spatial feature aggregation layer, is used to fuse the enhanced spatial feature representation with numerical weather forecast features, and output the power prediction value for the next 24 hours to 168 hours through a fully connected neural network. The optimization decision execution module is connected to the core power prediction module. It is used to construct an optimization model based on the power prediction value with the goal of minimizing the overall operating cost. After solving the model, it generates power control instructions for each new energy power station and sends the instructions to each power station for execution.

[0007] Preferably, the process of calculating dynamic attention weights in the spatial feature aggregation layer specifically includes: For target station i, its hidden state features are concatenated with the hidden state features of neighboring station j to form a combined feature vector; The combined feature vector is multiplied by a learnable attention vector, and the LeakyReLU nonlinear activation function is applied to obtain the unnormalized attention score. By performing softmax normalization on the attention scores obtained by all stations connected to station i, including the station itself, the dynamic attention weight coefficient α_{ij} of station j to station i is obtained. Based on the dynamic attention weight coefficient α_{ij}, the features of neighboring stations are weighted and summed to update the feature representation of target station i.

[0008] Preferably, the spatial feature aggregation layer employs a multi-head attention mechanism, specifically: K independent attention calculations are performed in parallel, where K is an integer greater than 1, and each calculation uses different learnable parameters; The updated feature representations obtained from K calculations are concatenated to form the final enhanced spatial feature representation for each station.

[0009] Preferably, a feature alignment layer is further included between the bidirectional temporal coding layer and the spatial feature aggregation layer, for: The features of each time step in the hidden state sequence are multiplied element-wise with the numerical weather forecast data at the corresponding time, so as to achieve preliminary alignment and filtering of time series features and meteorological features.

[0010] Preferably, the core power prediction module further includes an uncertainty quantization unit, used for: A probability prediction subnetwork is connected to the end of the bidirectional spatiotemporal neural network model. This subnetwork simultaneously outputs the conditional probability distribution parameters of the power prediction value, including the mean μ and the variance σ. When constructing the optimization model, the optimization decision execution module uses the variance σ as a measure of prediction uncertainty, calculates the prediction risk cost through a monotonically increasing penalty function, and incorporates this risk cost into the overall operating cost.

[0011] Preferably, the optimization decision execution module, when constructing the optimization model, is specifically used for: The objective function is to minimize the total operating cost, which includes: the cost of deviation from grid dispatch instructions, the cost of power adjustment at each power station, the cost of wind and solar curtailment, and the risk cost based on forecast uncertainty. The constraints include: upper and lower limits of power for each power station, power allocation ratio between power stations, power ramp rate constraints, and grid dispatch requirements. The optimal power allocation scheme is obtained by solving the optimization model using a constrained gradient descent algorithm or an interior point method.

[0012] Preferably, the system further includes an online model update module, used for: Real-time monitoring of the deviation between the actual output and the predicted value of each new energy power station; when the average absolute percentage error exceeds 5% for three consecutive cycles, the model incremental learning is triggered. Using the latest 30 days of data within the sliding time window, the parameters of some network layers of the bidirectional spatiotemporal neural network model are fine-tuned to adapt to changes in the characteristics of wind and solar resources.

[0013] Preferably, the data acquisition and preprocessing module further includes an abnormal data identification unit, used for: An anomaly detection algorithm based on isolated forests is used to identify and remove outliers in the power data; For missing data, time series decomposition and reconstruction methods are used for imputation to ensure data continuity.

[0014] Preferably, the system further includes a prediction result visualization module, used for: Power prediction results, optimization decision schemes, and key system performance indicators are visualized through a web interface; It provides a comparative analysis of power prediction curves and measured curves, as well as an evaluation report on the effectiveness of optimization decisions.

[0015] Preferably, the system supports centralized management of multiple new energy power stations, the number of which is between 5 and 50, and the types of power stations include at least one of wind power plants and photovoltaic power plants.

[0016] The technical effects and advantages of this invention are as follows: Compared to existing technologies, this invention improves prediction accuracy by constructing a core prediction model that combines a bidirectional gated recurrent unit (GRU) with a graph attention network. The GRU learns the temporal dependencies of power sequences from both positive and negative directions, capturing historical evolution patterns more comprehensively. The graph attention network, building on this, treats each station as a node in a graph, adaptively learning and aggregating spatial features of strongly correlated neighboring stations by calculating dynamic attention weights. This transforms spatiotemporal correlation information ignored by traditional methods into effective prediction features, enabling the model to more accurately characterize the coordinated fluctuation behavior of each station under the influence of the regional meteorological field. Ultimately, this reduces the prediction error of the overall power at the control center, providing a more reliable data foundation for optimization decisions.

[0017] This invention integrates a probabilistic prediction sub-network into its core prediction model, simultaneously outputting a variance parameter representing prediction uncertainty along with the predicted point values. The optimization decision module uses this variance as a key input, quantifying prediction uncertainty into a specific risk cost through a monotonically increasing penalty function, and incorporating it into an optimization model aimed at minimizing overall operating costs. This approach allows the decision-making process to proactively "sense" risk, automatically avoiding operating points with high prediction uncertainty when formulating power allocation schemes. Compared to traditional deterministic optimization, the decision schemes generated by this method exhibit stronger adaptability to actual power fluctuations, effectively reducing the risk of grid performance penalties due to prediction errors, and improving the system's economy and reliability.

[0018] This invention constructs a dynamic system capable of responding to environmental changes by introducing an online model update module and an anomaly data identification unit. The anomaly data identification unit automatically cleans input data based on the Isolation Forest algorithm, ensuring the quality of model input. The online model update module continuously monitors prediction performance and automatically triggers incremental learning when performance degrades, fine-tuning model parameters using recent data. Compared to a fixed model, this mechanism enables the system to continuously track long-term changes in wind and solar resource characteristics and the operational status of site equipment, avoiding the problem of continuous accuracy decline due to model "aging." This ensures the long-term effectiveness of the system's prediction and decision-making capabilities throughout its entire lifecycle and reduces the manpower and computational costs of frequent large-scale retraining in later stages. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall system workflow of the present invention.

[0020] Figure 2 This is a detailed flowchart of the core power prediction module of the present invention.

[0021] Figure 3 This is a diagram illustrating the spatial feature aggregation and attention mechanism of the present invention.

[0022] Figure 4 This is a diagram of the optimization decision-making and system support modules of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1 As attached Figures 1 to 4 The system shown is a power prediction and optimization decision-making system for a new energy centralized control center. The system consists of three core components that work together: The data acquisition and preprocessing module is responsible for integrating multi-source data and performing quality control. The core power prediction module generates high-precision prediction results based on an improved neural network architecture. The optimization decision execution module comprehensively considers economic efficiency and risk to formulate the optimal power allocation scheme.

[0025] These three modules are connected through standardized data interfaces, forming a closed-loop processing flow from data acquisition to decision execution. The system adopts a distributed architecture design, with front-end data processing nodes responsible for real-time acquisition of operational data from each site, intermediate computing nodes carrying the core algorithm model, and back-end management nodes responsible for result display and command issuance. The nodes exchange data through a high-speed network, ensuring that the system can maintain a high response speed when processing large-scale site cluster data.

[0026] Furthermore, a graph neural network-based processing flow is adopted when calculating dynamic attention weights in the spatial feature aggregation layer. This flow first constructs a power station association graph, where nodes represent individual renewable energy power stations and edges represent spatial relationships between stations. The graph construction considers not only geographical proximity but also multiple dimensions such as power station type, installed capacity, and historical output correlation. For the target power station... Hide its state features with neighboring stations Features The features are concatenated to form a combined feature vector. The combined features are computed through a learnable attention mechanism: in This represents the attention parameter vector. The final attention weight coefficients are obtained by normalizing the attention scores of all neighboring nodes. This computational approach enables the model to adaptively adjust the influence weights of neighboring wind farms on the target wind farm, thereby more accurately capturing the dynamic spatial correlation characteristics between wind farms. In actual calculations, the system calculates attention weights separately for each time step, reflecting the dynamic characteristics of the correlation between wind farms over time. For example, under weather conditions of rapidly changing wind speeds, the correlation weight between geographically proximate wind farms automatically increases, while the correlation weight between photovoltaic power plants and wind farms decreases accordingly. This dynamic adjustment mechanism enhances the model's ability to capture power fluctuation patterns under complex weather conditions.

[0027] Furthermore, the multi-head attention mechanism enhances the model's representational power by deploying eight independent attention heads in parallel. Each attention head uses a different set of parameters. ,in Calculate the corresponding attention weights for each attention head. The output features of each attention head are concatenated in the last dimension. This allows the model to capture information from different feature subspaces, effectively providing the system with multiple perspectives to analyze the complex inter-station correlation patterns. In actual operation, different attention heads spontaneously learn to focus on different types of spatial correlation patterns. For example, some attention heads specifically capture correlations arising from geographical proximity, while others focus on correlations generated by the similarity of station types, and still others learn correlations reflected in the similarity of historical output curves. This collaborative mechanism enables the model to more comprehensively understand the overall operational characteristics of the station cluster, especially demonstrating stronger adaptability when dealing with heterogeneous station clusters (including different types of new energy stations such as wind power and photovoltaics).

[0028] Furthermore, the feature alignment layer achieves deep fusion of temporal and meteorological features through a gating mechanism. This layer receives the hidden state sequence output by the bidirectional gated recurrent unit. and corresponding meteorological characteristics Feature alignment is achieved through projection transformation and element-wise multiplication: ,in This mechanism maps meteorological features to a learnable projection matrix of the same dimension as the temporal features. It dynamically adjusts the weights of temporal features based on their importance, increasing attention to these features when weather conditions change drastically. Specifically, when meteorological data indicates an impending change in cloud cover or a sudden shift in wind speed, the system automatically increases the weight coefficients of the meteorological features, enabling the model to respond more quickly to changes in the external environment.

[0029] Conversely, during periods of stable meteorological conditions, the system relies more heavily on the temporal patterns of historical power sequences for prediction. This adaptive weight adjustment mechanism effectively resolves the feature conflict problem caused by the simple splicing of multi-source features in traditional methods, and improves the model's ability to predict power fluctuations under complex meteorological conditions.

[0030] Furthermore, the uncertainty quantification unit evaluates the reliability of the prediction results by connecting a probability prediction subnetwork at the end of the model. This subnetwork contains two parallel fully connected layers, each outputting the mean of the predicted values. and variance The training process uses a negative log-likelihood loss function: This design enables the model to adaptively learn the level of uncertainty under different input conditions, providing important risk quantification basis for subsequent optimization decisions. Variance The calculation not only considers the statistical characteristics of historical prediction errors but also incorporates feature quality assessments of the current input data. For example, when the input meteorological data is of poor quality or incomplete, the model automatically increases the variance estimate, reflecting the low reliability of the current prediction results. This uncertainty quantification mechanism provides crucial risk perception capabilities for subsequent optimization decisions, enabling the system to distinguish between high-confidence and low-confidence predictions, thereby allowing for differentiated risk response strategies when making decisions.

[0031] Furthermore, the optimization decision-making module constructs an objective function that comprehensively considers both economic efficiency and risk: Each item corresponds to a scheduling deviation cost, power adjustment cost, wind and solar curtailment penalty cost, and risk cost, respectively. Power is measured in megawatts (MW), and cost is measured in yuan. Constraints include upper and lower power limits for each power station. Power allocation ratio constraints Slope rate constraint and power grid dispatching requirements and constraints .

[0032] This optimization problem is solved using a sequential quadratic programming algorithm, which iteratively solves a series of quadratic programming subproblems to gradually approach the optimal solution. During the solution process, the system prioritizes meeting grid dispatch requirements and equipment safety constraints, then seeks a balance between economic efficiency and risk. Risk preference coefficient. It can be dynamically adjusted according to the real-time operating status of the power grid. When the system's peak-shaving capacity is strained or the penetration rate of new energy sources is high, it can automatically increase the degree of risk avoidance to ensure the reliability of decision-making schemes.

[0033] Furthermore, the online model update module employs a sliding window incremental learning mechanism to maintain the model's timeliness. When predictive performance degradation is detected, the system automatically uses data from the most recent 30 days for incremental training. The update process keeps the parameters of the bidirectional temporal coding layer fixed, only fine-tuning the parameters of the spatial feature aggregation layer and the multi-source fusion prediction layer, with the learning rate set to 0.1, the same as in the initial training phase.

[0034] This partial fine-tuning strategy effectively prevents overfitting while ensuring the model adapts to new data distributions. The system triggers the update mechanism by continuously monitoring the statistical characteristics of the prediction error. When the mean absolute percentage error exceeds a set threshold of 5% for three consecutive prediction periods, the model update process is automatically initiated. During the update process, the system retains the general patterns learned by the model from long-term historical data and only adjusts parameters related to recent data distributions. This gradual update strategy ensures both the timeliness of the model and the stability of the prediction results.

[0035] Furthermore, the outlier identification unit combines the Isolation Forest algorithm and time series analysis methods for data quality control. The Isolation Forest algorithm identifies outliers by constructing randomly partitioned decision trees, and its outlier score is calculated based on the difficulty of isolating data points. For identified outliers, time series decomposition methods are used for repair. ; in For trend items, It is a periodic term. This represents the residual term. The trend and periodic terms are estimated using robust local regression, and then outliers are imputed based on the reconstructed sequence to ensure data continuity and reliability. The anomaly detection process consists of two stages: First, the Isolation Forest algorithm is used to quickly identify potential outliers, which are then confirmed using a time-series consistency check. For data points confirmed as anomalous, different handling strategies are adopted according to their degree of anomalousness: minor anomalies are repaired using neighbor interpolation, while severe anomalies trigger data quality alarms and notify operations personnel for intervention. This hierarchical processing mechanism ensures data quality while preserving the information content of the original data to the maximum extent.

[0036] Furthermore, the prediction results visualization module is built on a modern web technology stack to create an interactive analysis interface. The system provides a variety of visualization views, including a dual Y-axis line chart to show the comparison between predicted and measured values, a stacked area chart to show the power distribution of each site, a dashboard to show key performance indicators, and a network diagram to show the relationships between sites.

[0037] These visualization components acquire real-time data through a unified data interface, providing users with intuitive system operation status monitoring and decision analysis support. The visualization system adopts a component-based design, where each view can be independently configured and extended, allowing users to customize personalized monitoring interfaces according to their actual needs. The system also provides historical data backtracking functionality, allowing users to compare and analyze prediction performance over different time periods, providing data support for model optimization.

[0038] Example 2 In a specific application scenario of a provincial-level new energy centralized control center, the system manages 25 wind farms and 15 photovoltaic power stations, with a total installed capacity of 3.5GW.

[0039] The data acquisition server cluster continuously collects power data and equipment status data from various sites, while also obtaining numerical weather forecast data from meteorological service providers. Every day at 02:00, the system automatically starts the power prediction process, using historical data from the past 30 days to train the model and generate power prediction curves and uncertainty assessment results for the next 168 hours.

[0040] Based on these forecasts, the optimization decision-making module, combined with grid dispatch instructions and operational constraints, solves an optimization problem with risk constraints to generate the day-ahead power generation plan for each power station. During the intraday operation phase, the system performs rolling optimization every 15 minutes, dynamically adjusting power control instructions based on the latest ultra-short-term forecast results.

[0041] These instructions are issued to each renewable energy power station for execution through the automated scheduling system, while all key information is displayed to operators in real time through a visual interface. The entire system achieves intelligent management of the entire process from data acquisition to decision execution, providing solid technical support for the reliable grid connection and efficient consumption of large-scale renewable energy.

[0042] In the workflow of a specific operating day, the system initiates the daytime forecasting process at 02:00 AM. The data acquisition module first obtains the historical power output data of each station over the past 30 days, as well as the numerical weather forecast data for the next 7 days.

[0043] After data quality checks and outlier handling, the core prediction module begins operation. The bidirectional temporal coding layer first processes the power sequences of each wind farm, extracting temporal features through GRU networks in both forward and reverse directions. Assume the power sequence of a certain wind farm is... After bidirectional GRU processing, a hidden state sequence containing contextual information is obtained. .

[0044] Subsequently, the spatial feature aggregation layer begins to function. (Taking wind farms as an example.) For example, the system first determines its neighboring site set, including three adjacent wind farms and two photovoltaic power stations. For each neighboring site, it calculates its attention weight with the target site. Assume the wind farm... With wind farm Attention score is calculated as follows: Similarly, the attention scores of other neighboring stations are calculated, and the weight coefficients are obtained after softmax normalization: After eight attention heads are computed in parallel, the resulting feature representations are concatenated to form a wind farm. The final spatial characteristic representation. This characteristic representation has incorporated operational information from surrounding stations, and can better reflect the impact of regional meteorological conditions.

[0045] The feature alignment layer then fuses the temporal features with the meteorological features. Assuming that the current meteorological features indicate an impending increase in wind speed, the system uses a gating mechanism to enhance the weight of the meteorological features, enabling the prediction model to respond to this change in a timely manner.

[0046] The multi-source fusion prediction layer receives the processed features and outputs power predictions for the next 168 hours. The uncertainty quantization unit simultaneously outputs the prediction variance for each time point. Near weather turning points, the variance value automatically increases, reflecting information that the prediction reliability is lower during these periods.

[0047] The optimization decision-making module formulates a power generation plan based on the forecast results. When solving the optimization problem, the system first addresses hard constraints to ensure that the power output of each power station does not exceed its limits and that the total output meets dispatch requirements. Then, within this feasible solution space, it seeks the optimal balance between economic efficiency and risk. For periods with high forecast uncertainty, the system appropriately increases the risk aversion coefficient and selects a more conservative operating strategy.

[0048] During the day's operation, the system performs rolling optimization every 15 minutes. When actual weather conditions deviate from the forecast, the dynamic feedback correction mechanism automatically triggers an update to the prediction model and adjusts the power allocation scheme for subsequent periods accordingly. This closed-loop control mechanism ensures that the system maintains excellent performance throughout the entire operating cycle.

[0049] Through the complete workflow described above, the system achieves fully automated processing across the entire chain, from data acquisition and feature extraction to power prediction and optimization decision-making, providing reliable technical support for the daily operation of the new energy control center. The close collaboration between the various modules ensures that the system maintains stable performance output and reliable decision-making quality even in the face of complex and ever-changing operating environments.

[0050] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power prediction and optimization decision-making system for a new energy centralized control center, characterized in that, include: The data acquisition and preprocessing module is used to acquire historical power data, real-time operation data and numerical weather forecast data of each new energy power station under the central control center, and to perform quality verification and normalization processing on the data to form a standardized training sample set. The core power prediction module, connected to the data acquisition and preprocessing module, integrates a bidirectional spatiotemporal neural network model based on an attention mechanism. This model processes the training sample set and outputs power prediction values ​​for a specific future time period. The bidirectional spatiotemporal neural network model sequentially includes: The bidirectional temporal coding layer employs a bidirectional gated recurrent unit network to extract the temporal features of the power sequence of each new energy power station in both forward and reverse order, and outputs a hidden state sequence containing contextual information. The spatial feature aggregation layer, connected to the bidirectional temporal coding layer, adopts a graph attention network. By calculating the dynamic attention weights between any two stations, it aggregates the spatial features of adjacent stations and generates an enhanced spatial feature representation for each station. The multi-source fusion prediction layer, connected to the spatial feature aggregation layer, is used to fuse the enhanced spatial feature representation with numerical weather forecast features, and output the power prediction value for the next 24 hours to 168 hours through a fully connected neural network. The optimization decision execution module is connected to the core power prediction module. It is used to construct an optimization model based on the power prediction value with the goal of minimizing the overall operating cost. After solving the model, it generates power control instructions for each new energy power station and sends the instructions to each power station for execution.

2. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The process of calculating dynamic attention weights in the spatial feature aggregation layer specifically includes: For target station i, its hidden state features are concatenated with the hidden state features of neighboring station j to form a combined feature vector; The combined feature vector is multiplied by a learnable attention vector, and the LeakyReLU nonlinear activation function is applied to obtain the unnormalized attention score. By performing softmax normalization on the attention scores obtained by all stations connected to station i, including the station itself, the dynamic attention weight coefficient α_{ij} of station j to station i is obtained. Based on the dynamic attention weight coefficient α_{ij}, the features of neighboring stations are weighted and summed to update the feature representation of target station i.

3. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 2, characterized in that, The spatial feature aggregation layer employs a multi-head attention mechanism, specifically: K independent attention calculations are performed in parallel, where K is an integer greater than 1, and each calculation uses different learnable parameters; The updated feature representations obtained from K calculations are concatenated to form the final enhanced spatial feature representation for each station.

4. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, Between the bidirectional temporal coding layer and the spatial feature aggregation layer, a feature alignment layer is also included, used for: The features of each time step in the hidden state sequence are multiplied element-wise with the numerical weather forecast data at the corresponding time, so as to achieve preliminary alignment and filtering of time series features and meteorological features.

5. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The core power prediction module also includes an uncertainty quantization unit, used for: A probability prediction subnetwork is connected to the end of the bidirectional spatiotemporal neural network model. This subnetwork simultaneously outputs the conditional probability distribution parameters of the power prediction value, including the mean μ and the variance σ. When constructing the optimization model, the optimization decision execution module uses the variance σ as a measure of prediction uncertainty, calculates the prediction risk cost through a monotonically increasing penalty function, and incorporates this risk cost into the overall operating cost.

6. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The optimization decision execution module, when constructing the optimization model, is specifically used for: The objective function is to minimize the total operating cost, which includes: the cost of deviation from grid dispatch instructions, the cost of power adjustment at each power station, the cost of wind and solar curtailment, and the risk cost based on forecast uncertainty. The constraints include: upper and lower limits of power for each power station, power allocation ratio between power stations, power ramp rate constraints, and grid dispatch requirements. The optimal power allocation scheme is obtained by solving the optimization model using a constrained gradient descent algorithm or an interior point method.

7. The power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The system also includes an online model update module, used for: Real-time monitoring of the deviation between the actual output and the predicted value of each new energy power station; when the average absolute percentage error exceeds 5% for three consecutive cycles, the model incremental learning is triggered. Using the latest 30 days of data within the sliding time window, the parameters of some network layers of the bidirectional spatiotemporal neural network model are fine-tuned to adapt to changes in the characteristics of wind and solar resources.

8. A power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The data acquisition and preprocessing module also includes an abnormal data identification unit, used for: An anomaly detection algorithm based on isolated forests is used to identify and remove outliers in the power data; For missing data, time series decomposition and reconstruction methods are used for imputation to ensure data continuity.

9. A power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The system also includes a prediction result visualization module for: Power prediction results, optimization decision schemes, and key system performance indicators are visualized through a web interface; It provides a comparative analysis of power prediction curves and measured curves, as well as an evaluation report on the effectiveness of optimization decisions.

10. A power prediction and optimization decision-making system for a new energy centralized control center according to claim 1, characterized in that, The system supports centralized management of multiple new energy power stations, the number of which ranges from 5 to 50, and the station types include at least one of wind power plants and photovoltaic power plants.

Images