A power system optimization dispatching method based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based power system optimization and scheduling method, which relates to the technical field of power system fault testing. The method comprises the following steps: obtaining production monitoring data, operation data and management data under the target power system as a power data set, and performing preprocessing; constructing a load forecasting model, inputting historical load data, user electricity consumption behavior, meteorological data and calendar characteristics, and outputting a predicted power grid load curve. The technical key points are: by combining the equipment health status assessment model, dynamically adjusting the health score and subsequent real-time scheduling optimization, a highly interconnected optimization management plan is formed, which achieves the effect of accurately identifying potential fault risks, enables operation and maintenance personnel to take preventive measures in advance, avoids misjudgment or omission due to ignoring external factors, and enhances the prediction and response capabilities of the entire scheduling plan to equipment health risks.

Description

Technical Field

[0001] The present invention relates to the technical field of power system fault testing, and in particular to an artificial intelligence-based power system optimization scheduling method. Background Art

[0002] At a time when digital technology and industry are rapidly converging, data has become a new production factor, and artificial intelligence has become a new production tool. Artificial intelligence will be used to assist in tasks such as power system fault testing and optimized scheduling. Data analysis of parameters in the power system is carried out. The data analysis here is to analyze the original data and search for specific types of patterns and trends. The analysis results are used to define the optimal parameters for creating a mining model, and these parameters are applied to the entire data set to extract feasible patterns and detailed statistical information to obtain more effective information in order to understand historical situations, monitor current conditions, speculate and analyze causes, predict future developments, or assist in decision-making.

[0003] At present, traditional power system equipment health management mostly relies on static scoring models or simple judgments based on thresholds, and lacks an effective response mechanism to dynamic changes in the operating environment, resulting in delayed health scoring and high misjudgment rate. For example, in the scenario of sudden load increase, traditional methods are difficult to timely identify the deterioration trend of equipment health status, resulting in delayed maintenance or even sudden failures, affecting power supply reliability. At the same time, existing solutions generally do not include load change rate as a key parameter in scoring modeling. Even if it is partially considered, it is impossible to achieve dynamic matching between the model and the operating scenario, resulting in poor scoring sensitivity and difficulty in providing effective support for scheduling decisions. In addition, most systems lack a closed-loop feedback mechanism, and the model parameters are fixed for a long time or the update steps are accurate. It is impossible to learn and optimize from the actual operating results, resulting in persistent scoring deviations. This makes the operation and maintenance strategies often rely on experience judgment, unreasonable resource allocation, high maintenance costs and insufficient system stability, which seriously affects the intelligent management level of the power system. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0005] An artificial intelligence-based power system optimization and dispatching method includes the following steps: obtaining production monitoring data, operation data, and management data of a target power system as a power data set, and preprocessing the data;

[0006] Construct a load forecasting model, input historical load data, user electricity consumption behavior, meteorological data, and calendar characteristics, and output a predicted grid load curve. Simultaneously extract key performance indicators from the equipment's operation log and combine them with equipment ledger information as static features. Use supervised learning to model the health score and derive the original health score H0. Introduce a load factor L based on the load change rate to derive the final health score H, where L = f().

[0007] Construct a dispatch optimization objective function, which at least includes: minimizing power generation costs, balancing loads, and the N-1 safety criterion. After constraint modeling, start the learning optimization algorithm and output the dispatch execution plan. Based on the dispatch execution results, calculate the equipment health score deviation between the simulation and the actual dispatch execution. When the health score deviation exceeds the set indicator, perform parameter correction actions based on the health score deviation; otherwise, no response action is taken.

[0008] Furthermore, production monitoring data includes at least: power generation, voltage and current, spare capacity, equipment inventory information and operation logs; operation data includes at least: electricity prices, user electricity usage behavior, electricity sales and load curves; management data includes at least: staffing, project progress, maintenance resources and cost data.

[0009] Furthermore, the preprocessing of the power data set includes: using a machine learning anomaly detection algorithm to identify and repair abnormal data in the power data set, converting the power data set into a unified data structure, and performing time series synchronization on the power data set.

[0010] Furthermore, the meteorological data includes at least: temperature and humidity, wind speed and light intensity, and the calendar characteristics include at least: holidays and seasons;

[0011] When building the load forecasting model, either LSTM or Transformer network is used; the sliding window method is used to construct training samples, and the data input to the load forecasting model is divided into training set, validation set and test set.

[0012] Furthermore, the key performance indicators extracted from the equipment operation log include at least: temperature, vibration, oil pressure, insulation resistance and load rate; the combined equipment ledger information includes at least: model, manufacturing year and maintenance times;

[0013] When modeling health score, feature extraction also includes: marking whether a fault has occurred and obtaining a fault label;

[0014] Among them, 0 indicates normal and 1 indicates fault.

[0015] Furthermore, the process of constructing a device health score using supervised learning is as follows:

[0016] The classification model is trained with the fault label as the dependent variable and the key performance indicators and static features as the independent variables. The classification model outputs: the probability of failure of the device at a certain point in the future. The predicted failure probability is mapped to a health score.

[0017] The formula is: original health score H0=100*P fault Among them, P fault represents the probability of failure.

[0018] Furthermore, the basis for obtaining the final health score H is as follows:

[0019] The health score H is expressed as a combination of the original health score H0 and the load factor L: H = H0 + α * L; where α is the weight, ranging from 0 to 1; and the setting process of the function f is as follows:

[0020] Depending on the constraints:

[0021] Condition 1: The load change rate is within the standard threshold {Thd1, Thd2};

[0022] Condition 2: The load change rate is higher than Thd2;

[0023] Condition 3: The load change rate is lower than Thd1;

[0024] When condition 1 is met, the function f adopts the basic model:

[0025] ;

[0026] Where fl represents future load, cl represents current load, ld represents load, and t represents unit duration; Indicates the load change rate; w1 and w2 represent weight coefficients, and their value ranges are both greater than 0;

[0027] When the second condition is met, the function f adopts a nonlinear enhancement model:

[0028] ;

[0029] When condition three is met, the function f adopts a smoothing model:

[0030] .

[0031] Furthermore, the content of constraint modeling includes at least: generator start and stop constraints: each generator is set with start time and stop time requirements; transmission line capacity constraints: each transmission line is equipped with a maximum transmission power limit; equipment health status warning constraints: when the health score of the equipment exceeds the set threshold, an early warning is triggered.

[0032] Furthermore, the learning optimization algorithms launched include: Deep Q-Network and Genetic + Neural Network Hybrid Optimization Algorithm;

[0033] The content of the scheduling execution plan shall at least include:

[0034] Generate unit start and stop plans: adjust the start and stop plans of the generator units based on the current load demand and equipment health status; load distribution plan: distribute the load to each area to balance the load; equipment scheduling strategy: the final health score H is inversely proportional to the load capacity. When the health score H of the equipment exceeds the set threshold multiple Q, the equipment is suspended. If there is a backup equipment, it is activated; the value range of Q is greater than 1.

[0035] Furthermore, the health score deviation is: ΔH = H_1 - H_2; where H_1 represents the health score of the device in the actual scheduling execution, and H_2 represents the health score of the device in the simulated scheduling execution. The executed parameter correction action is used to feedback and correct the weight coefficients in the target model used in the corresponding function f, as follows:

[0036] ;

[0037] Where η represents the learning rate, and the value range of η is less than 0.01.

[0038] The present invention provides an artificial intelligence-based power system optimization scheduling method, which has the following beneficial effects:

[0039] (1) By combining the equipment health status assessment model, dynamically adjusting the health score, and subsequent real-time scheduling optimization, a highly interconnected optimization management solution is formed, which achieves the effect of accurately identifying potential failure risks, allowing operation and maintenance personnel to take preventive measures in advance to avoid misjudgment or omission due to ignoring external factors. This solves the problem that in traditional response solutions, due to the lack of an effective dynamic adjustment mechanism, it is often difficult to accurately predict equipment failures, resulting in increased maintenance costs and reduced system stability;

[0040] (2) In this scheme, the load change rate is introduced as a key parameter;

[0041] On the one hand, it can reflect the stability and volatility of the power system's operating status, and divide the system's operating status into three types of constraints based on the load change rate. By introducing a constraint classification mechanism, it achieves a precise match between the function model and the actual operating scenario, avoiding the errors caused by the use of a single fixed function, and thus improving the stability and credibility of subsequent health scoring schemes during operation. On the other hand, the load change rate is used as a function input parameter to enhance the sensitivity and accuracy of load factor calculations, enabling the health score to reflect the direct impact of load changes on the health status of the equipment in real time, improving the sensitivity and predictive ability of the score, and solving the technical problem that traditional static scoring models cannot provide timely warnings of equipment health deterioration.

[0042] (3) This scheme combines the load change rate with supervised learning and dynamic function modeling to realize the health score generation mechanism under multi-dimensional information fusion. It can achieve early warning, optimal resource allocation, and assist scheduling decisions through health scores, which helps to reduce power outages caused by sudden failures.

[0043] (4) Through the use of the feedback correction mechanism, on the one hand, the function model has the ability to learn and can gradually approach the real equipment response law, reflecting the adaptability of the AI ​​system; on the other hand, a more accurate load factor L helps to correct the health score model, avoid underestimating equipment risks due to ignoring external load pressure, and enhance the entire scheduling plan's ability to predict and respond to equipment health risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The figure is a schematic diagram of the overall process of an artificial intelligence-based power system optimization scheduling method in the present invention. DETAILED DESCRIPTION

[0045] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] Example 1:

[0047] See also Figure 1 This embodiment provides an AI-based method for optimizing power system scheduling. This method implements a closed-loop process based on "perception-analysis-decision-feedback," integrating three types of power system big data (corresponding to production monitoring, operations, and management, as discussed later). This method uses AI modeling to implement load forecasting, equipment health assessment, and scheduling optimization. The specific steps of this method are described below:

[0048] S1. Data collection and integration (perception layer):

[0049] S1.1. Electricity Dataset Collection:

[0050] Acquire production monitoring data, operation data, and management data of the target power system;

[0051] Production monitoring data includes: power generation, voltage and current, spare capacity, equipment inventory information, and operation logs; sources include: configured SCADA systems, PMU devices, and smart meters; operational data includes: electricity prices, user electricity usage behavior, electricity sales, and load curves; sources include: configured marketing systems and market trading platforms; management data includes: staffing, project progress, maintenance resources, and cost data; sources include: configured ERP systems, collaborative office platforms, and material management systems;

[0052] S1.2. Data cleaning and standardization:

[0053] Use machine learning anomaly detection algorithms (such as Isolation Forest) to identify and correct anomalies in the power dataset collected in S1.1. Convert power datasets from different sources or systems into a unified data structure for easy subsequent processing. This unified data structure means converting power datasets from different sources or systems into a standardized form with consistent format, type, or units to ensure data compatibility. Synchronize the power datasets to eliminate sampling frequency differences.

[0054] S1.3. Establish a unified data set:

[0055] Use distributed storage and computing platforms such as Hadoop, Spark, and Flink to build power data sets (i.e., power data sets include: production monitoring data, operation data, and management data); achieve centralized storage and rapid retrieval of multi-dimensional data, and support subsequent AI model training and real-time analysis.

[0056] S2. Load forecasting and equipment health status modeling (analysis layer):

[0057] S2.1. Build a load forecasting model based on LSTM / Transformer:

[0058] S2.1.1. Model input features:

[0059] Historical load data, user electricity usage behavior, meteorological data, and calendar features. Among the various data types used as model input features, historical load data and user electricity usage behavior (i.e., clustered load curves) are derived from the power dataset in S1. Meteorological data includes temperature, humidity, wind speed, and light intensity, and calendar features include holidays and seasons. Meteorological data and calendar features can be obtained online using IoT technology.

[0060] S2.1.2. Model selection and training:

[0061] LSTM or Transformer networks are used to capture long-term dependencies. The sliding window method is used to construct training samples, which are the input data of the load forecasting model. The input data of the load forecasting model is divided into training set, validation set, and test set. The prediction accuracy is evaluated using relevant indicators such as RMSE.

[0062] S2.1.3. Model result output:

[0063] Predicting future grid load curves (i.e., the basis for future loads); the future time range can be selected as needed. In this embodiment, 24 hours to 7 days (i.e., 168 hours) can be selected to achieve hourly prediction. In addition, it can also support power generation planning, peak load strategy optimization, and demand response arrangements, which will not be detailed here;

[0064] S2.2. Equipment health status assessment model (PHM model);

[0065] S2.2.1, Feature extraction:

[0066] Extract key performance indicators (KPIs) from the equipment's operation logs (equipment used in the target power system): temperature, vibration, oil pressure, insulation resistance, and load factor. Combined with equipment log information (model, year of manufacture, and number of maintenance visits), these KPIs are used as static features to identify whether a fault has occurred, yielding a fault label (0 for normal, 1 for fault).

[0067] S2.2.2. Health score modeling:

[0068] Use supervised learning (such as XGBoost or LightGBM) or unsupervised learning (such as Autoencoder) to model the health score and obtain the original health score H0. Then, introduce the load factor L and dynamically adjust the score according to the results of different stages of the grid load forecast to obtain the final health score H. L = f(), where f is a function that can be designed according to the specific situation.

[0069] The process of constructing a device health score using supervised learning in this embodiment is as follows:

[0070] The classification model is trained with the fault label as the dependent variable and the KPIs and static features as the independent variables. The classification model outputs the probability of failure of the device at a certain point in the future. The predicted failure probability is mapped to a health score, that is, the probability of 0 to 1 is converted to a score of 0 to 100. The formula is: original health score H0 = 100 * P fault , where P fault Represents the probability of failure. For example, if the model predicts that the failure probability of a certain device within the next week is 0.7 based on its input features, its health score is 70, indicating that it needs special attention and maintenance arrangements. This method uses historical data to train the model, achieving a mapping from multidimensional features to health scores, and is highly interpretable and practical.

[0071] The adjusted health score is based on the following:

[0072] The health score H is expressed as a combination of the original health score H0 and the load factor L: H = H0 + α * L, where α is a weight used to adjust the influence of the load factor and ranges from 0 to 1. In this embodiment, the value is usually 1.

[0073] In this embodiment, the function f is designed as follows:

[0074] According to different constraints (involving load impact analysis):

[0075] Condition 1: The load change rate is within the standard threshold {Thd1, Thd2};

[0076] Condition 2: The load change rate is higher than Thd2;

[0077] Condition 3: The load change rate is lower than Thd1;

[0078] When condition 1 is met, the function f adopts the basic model:

[0079] ;

[0080] Where fl represents future load, cl represents current load, ld represents load, and t represents unit duration;

[0081] Indicates the load change rate (the amount of change per unit time);

[0082] w1 and w2 represent weight coefficients, with a value range greater than 0. Each weight coefficient can be adjusted according to specific circumstances, or determined using the coefficient of variation method. It should be noted that the coefficient of variation method is a method of assigning weights to each indicator based on the degree of variation between the current value and the target value of each evaluation indicator. If the numerical difference of an indicator is large, it can clearly distinguish the evaluated objects, indicating that the indicator has rich discrimination information, and thus the indicator should be given a larger weight. Conversely, if the numerical difference of each evaluated object on a certain indicator is small, then the ability of this indicator to distinguish the evaluated objects is weak, and thus the indicator should be given a smaller weight. This method directly uses the information contained in each indicator to obtain the weight of the indicator through calculation, and therefore is objective.

[0083] The logic description of the basic model above is as follows:

[0084] When the load change rate is high, w2 should be large; if the load changes smoothly, it mainly depends on ;

[0085] When the second condition is met, the function f adopts a nonlinear enhancement model:

[0086] ;

[0087] In the formula, the exponential function is used To amplify the impact of high load change rate, e is the base, that is, the natural constant; is the index;

[0088] The logic is as follows: the exponential function significantly amplifies its impact under high load change rates, causing the health score to rise rapidly. This method is particularly suitable for scenarios that require a strong response to high load change rates.

[0089] When condition three is met, the function f adopts a smoothing model:

[0090] ;

[0091] In the formula, the logarithmic function log is used to smooth the impact of low load change rate;

[0092] The logical explanation is: the logarithmic function will weaken its impact at low load change rates, making the health score change more gradually; this method is suitable for situations where the load changes are relatively stable and avoids excessive sensitivity.

[0093] The reason for the above design to dynamically adjust the scoring method is:

[0094] The first is to improve accuracy: by considering real-time load conditions, the health score can be made more in line with the actual operating environment, avoiding misjudgment due to ignoring external factors; in fact, it is an early warning: when the load increases, even if there is no obvious abnormality in the equipment at present, the health score can be adjusted to remind the operation and maintenance personnel to pay attention to potential risks, so that preventive measures can be taken; finally, resource allocation is optimized: based on the dynamically adjusted health score, maintenance plans can be better arranged to ensure that sufficient resources are invested when they are most needed, reducing unnecessary downtime and costs; this design not only improves the accuracy of equipment health management, but also enhances the flexibility and response speed of the entire power system, which helps to achieve more intelligent operation and maintenance management; in summary, the higher the health score H, the worse the health status of the equipment, and the two are inversely proportional; and the equipment is related equipment used in the target power system.

[0095] By combining an equipment health assessment model, dynamically adjusting health scores, and subsequent real-time scheduling optimization, this solution forms a highly interconnected optimization management plan that accurately identifies potential failure risks. This approach allows operations and maintenance personnel to take preventative measures in advance, avoiding misjudgments or missed reports caused by ignoring external factors. This addresses the problem in traditional technical solutions, which often lack an effective dynamic adjustment mechanism and make it difficult to accurately predict equipment failures, leading to increased maintenance costs and decreased system stability.

[0096] Among them, for the load change rate;

[0097] On the one hand, it reflects the stability and volatility of the power system's operating state and is a key parameter for determining whether the equipment's operating environment is "harsh." In this solution, the system's operating state is divided into three types of constraints based on the load change rate. By introducing this constraint classification mechanism, a precise match between the function model and the actual operating scenario is achieved, avoiding the errors caused by the use of a single fixed function, and thus improving the stability and credibility of the subsequent health scoring scheme during operation.

[0098] Furthermore, using the load change rate as a function input parameter enhances the sensitivity and accuracy of load factor calculations. This design enables the health score to reflect the direct impact of load changes on equipment health in real time, improving the score's sensitivity and predictive capabilities. In particular, in the event of a sudden increase or decrease in load, potential risks can be quickly identified and the health score can be adjusted to alert operations and maintenance personnel to take preventive measures. This addresses the technical challenge of traditional static scoring models, which cannot provide timely warnings of equipment health deterioration.

[0099] In summary, in power systems, the health status of equipment is not only affected by its own degree of aging, but is also deeply constrained by the external operating environment (such as load changes). Therefore, relying solely on static characteristics or historical fault data for health scoring often fails to fully reflect the true status of the equipment. This solution introduces the load change rate as a key variable, combines supervised learning with dynamic function modeling, and implements a health score generation mechanism based on multi-dimensional information fusion. This can achieve early warning, optimal resource allocation, and assist scheduling decisions through health scores, helping to reduce power outages caused by sudden failures.

[0100] S3. AI-driven scheduling optimization decision-making (decision-making layer):

[0101] S3.1. Constructing a scheduling optimization objective function (aiming to balance multiple key indicators):

[0102] Minimize generation costs, balance loads, and meet the N-1 safety criteria;

[0103] It should be noted that minimizing power generation costs means optimizing the start and stop schedules of generator sets to select the power generation combination that minimizes costs; balancing loads means distributing loads across regions to avoid local overloads or underloads; and the N-1 safety principle ensures that the entire power system can continue to operate normally even in the event of a failure of a key component, thereby ensuring power supply reliability.

[0104] S3.2, Constraint modeling (considering multiple constraints):

[0105] Generator start and stop constraints: Each generator is set with start and stop time requirements to avoid frequent starts and stops; Transmission line capacity constraints: Each transmission line has a maximum transmission power limit to avoid exceeding its design capacity; Equipment health status warning constraints: When the equipment health score exceeds the set threshold, an alert is triggered;

[0106] S3.3. Start the learning optimization algorithm:

[0107] The learning optimization algorithms used include: Deep Q Network and Genetic + Neural Network Hybrid Optimization Algorithm;

[0108] Among them, the deep Q network: by simulating the decision-making process of dispatchers in complex environments, gradually learns the optimal strategy; the genetic + neural network hybrid optimization algorithm: by combining or combining the global search capabilities of the genetic algorithm and the powerful fitting capabilities of the neural network, it finds the global optimal solution;

[0109] S3.4, Scheduling output:

[0110] The scheduling execution output is based on a learning optimization algorithm and includes:

[0111] Generate unit start and stop plans: Adjust the start and stop plans of generator units based on current load demand and equipment health status; Load distribution plan: Allocate load to each area to balance the load; Equipment scheduling strategy: The lower the equipment health score, the more load it will be assigned, and the health score is inversely proportional to the load capacity; When the health score of a device exceeds the set threshold multiple Q, the device is suspended, and if there is a backup device, it is activated;

[0112] Where Q>1, for example, 2 times the set threshold. If the health score of a device exceeds 2 times the set threshold, it indicates that the device is faulty or close to faulty and needs to be shut down for maintenance.

[0113] The above-mentioned scheme adopts the technical solution of load forecasting model combined with reinforcement learning or hybrid optimization algorithm to achieve the ability to respond to emergencies efficiently. This enables the power system to make rapid adjustments in the face of extreme weather or other unforeseen events, and solves the problems that traditional scheduling optimization methods only rely on fixed rules, have poor flexibility, are difficult to respond quickly and effectively when encountering emergencies, and are prone to causing large-scale power outages. It realizes and ensures the reliability of power supply.

[0114] S3.5 Feedback Correction Mechanism (Feedback Layer):

[0115] Based on the scheduling execution results, the deviation of the equipment health score between the simulated and actual scheduling execution is calculated: ΔH = H_1 - H_2, where ΔH represents the equipment health score deviation, H_1 represents the equipment health score of the actual scheduling execution, and H_2 represents the equipment health score of the simulated scheduling execution. When ΔH exceeds the set indicator (indicating that the deviation is outside the acceptable range), parameter correction actions are performed based on the health score deviation, which is used to feedback and correct the weight coefficients in the target model used by function f in S2.2.2. The content is as follows:

[0116] ;

[0117] Where η is the learning rate, which controls the magnitude of each correction (usually set to 0.01 or less).

[0118] The two results obtained from the above calculation, namely w1' and w2', represent the modified weight coefficients;

[0119] Logical explanation: Health score deviation directly reflects the actual impact of scheduling behavior on equipment health. If there is a significant deviation between the simulated health score and the actual result, it means that the load factor function model fails to accurately reflect the equipment response patterns in the real environment. By using this deviation to make reverse corrections, the function model weights can be dynamically adjusted to better adapt to different equipment types, operating environments, and load change patterns.

[0120] Otherwise, no response action will be taken;

[0121] In summary, the entire S3 solution, through training in a simulation environment and optimization using scheduling execution results, can provide more efficient guidance for subsequent warning triggering;

[0122] Effect description: Through the use of the feedback correction mechanism, on the one hand, the function model has the ability to learn and can gradually approach the real equipment response law, reflecting the adaptability of the AI ​​system; on the other hand, the more accurate load factor L helps to correct the health score model, avoid underestimating equipment risks due to ignoring external load pressure, and enhance the entire scheduling plan's ability to predict and respond to equipment health risks.

[0123] Example 2:

[0124] Based on Example 1, this embodiment further provides an artificial intelligence-based power system optimization and dispatching system, which includes functional modules or units that operate in sequence, as follows:

[0125] Sensing and acquisition unit: acquires production monitoring data, operation data, and management data of the target power system as a power data set and performs preprocessing;

[0126] Analysis and Management Unit: Builds a load forecasting model, inputs historical load data, user electricity consumption behavior, meteorological data, and calendar characteristics, and outputs a predicted grid load curve. Key performance indicators are simultaneously extracted from the equipment's operation logs and combined with equipment ledger information as static features. Supervised learning is used to model health scores and derive the original health score H0. A load factor L is introduced based on the load change rate to derive the final health score H, where L = f(current load, future load).

[0127] Decision feedback unit: Constructs a scheduling optimization objective function, which at least includes: minimizing power generation costs, balancing loads, and the N-1 safety criterion. After constraint modeling, it starts the learning optimization algorithm and outputs a scheduling execution plan. Based on the scheduling execution results, it calculates the equipment health score deviation between the simulation and the actual scheduling execution. When the health score deviation exceeds the set indicator, it performs parameter correction actions based on the health score deviation; otherwise, it does not take any response action.

[0128] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0129] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0130] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A power system optimization dispatching method based on artificial intelligence, characterized in that: The method includes the following steps: obtaining production monitoring data, operation data and management data of a target power system as a power data set, and performing preprocessing; Build a load forecasting model, input historical load data, user electricity consumption behavior, meteorological data and calendar characteristics, and output the predicted grid load curve; Key performance indicators are extracted from the equipment's operation logs and combined with equipment ledger information as static features. Supervised learning is used to model the health score and obtain the original health score H0. The load factor L is introduced according to the load change rate to obtain the final health score H; where L = f(); f is a function; The final health score H is derived based on the following: The health score H is expressed as a combination of the original health score H0 and the load factor L: H = H0 + α*L; where α is a weight ranging from 0 to 1; and the setting process of the function f is as follows: Depending on the constraints: Condition 1: The load change rate is within the standard threshold {Thd1, Thd2}; Condition 2: The load change rate is higher than Thd2; Condition 3: The load change rate is lower than Thd1; When condition 1 is met, the function f adopts the basic model: L=w1·(fl-cl)+w2·Δld / Δt; Where fl represents future load, cl represents current load, ld represents load, and t represents unit duration; Δld / Δt represents load change rate; w1 and w2 represent weight coefficients, both of which have a value range greater than 0; When the second condition is met, the function f adopts a nonlinear enhancement model: L=w1·(fl-cl)+w2·(e ΔldΔt -1); Where, e represents a natural constant; When condition three is met, the function f adopts a smoothing model: L=w1·(fl-cl)+w2·log(1+Δld / Δt); Construct a dispatch optimization objective function, which at least includes: minimizing power generation costs, balancing loads, and the N-1 safety criterion. After constraint modeling, start the learning optimization algorithm and output the dispatch execution plan. Based on the dispatch execution results, calculate the equipment health score deviation between the simulation and the actual dispatch execution. When the health score deviation exceeds the set indicator, perform parameter correction actions based on the health score deviation; otherwise, no response action is taken.

2. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: Production monitoring data shall at least include: power generation, voltage and current, spare capacity, equipment inventory information and operation logs; operation data shall at least include: electricity prices, user electricity usage behavior, electricity sales situation and load curves; management data shall at least include: staffing, project progress, maintenance resources and cost data.

3. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: The preprocessing of the power data set includes: using machine learning anomaly detection algorithms to identify and repair abnormal data in the power data set, converting the power data set into a unified data structure, and performing time series synchronization on the power data set.

4. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: Meteorological data shall at least include: temperature, humidity, wind speed and light intensity; calendar characteristics shall at least include: holidays and seasons; When building the load forecasting model, either LSTM or Transformer network is used; the sliding window method is used to construct training samples, and the data input to the load forecasting model is divided into training set, validation set and test set.

5. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: Extract key performance indicators from the equipment's operation log, including at least temperature, vibration, oil pressure, insulation resistance, and load factor; combine with equipment ledger information, including at least model, year of manufacture, and number of maintenance visits; When modeling health score, feature extraction also includes: marking whether a fault has occurred and obtaining a fault label; Among them, 0 indicates normal and 1 indicates fault.

6. The power system optimization and dispatching method based on artificial intelligence according to claim 5, characterized in that: The process of constructing a device health score using supervised learning is as follows: The classification model is trained with the fault label as the dependent variable and the key performance indicators and static features as the independent variables. The classification model outputs: the probability of failure of the device at a certain point in the future. The predicted failure probability is mapped to a health score. The formula is: original health score H0 = 100*Pfault; where Pfault represents the probability of failure.

7. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: The content of constraint modeling includes at least: generator start and stop constraints: each generator has start and stop time requirements; transmission line capacity constraints: each transmission line is equipped with a maximum transmission power limit; equipment health status warning constraints: when the health score of the equipment exceeds the set threshold, an early warning is triggered.

8. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: The learning optimization algorithms launched include: Deep Q Network and Genetic + Neural Network Hybrid Optimization Algorithm; The content of the scheduling execution plan shall at least include: Generate unit start and stop plans: adjust the start and stop plans of the generator units based on the current load demand and equipment health status; load distribution plan: distribute the load to each area to balance the load; equipment scheduling strategy: the final health score H is inversely proportional to the load capacity. When the health score H of the equipment exceeds the set threshold multiple Q, the equipment is suspended. If there is a backup equipment, it is activated; the value range of Q is greater than 1.

9. The power system optimization and dispatching method based on artificial intelligence according to claim 1, characterized in that: Health score deviation: ΔH = H_1 - H_2; where H_1 represents the health score of the device in the actual scheduling execution, and H_2 represents the health score of the device in the simulated scheduling execution. The executed parameter correction action is used to feedback and correct the weight coefficients in the target model used in the corresponding function f, as follows: Where η represents the learning rate, and the value range of η is less than 0.01; w1' and w2' represent the modified weight coefficients.

Images