Artificial intelligence observation station network layout optimization method combining PID and feature contribution analysis

Abstract

The embodiment of the present disclosure provides an artificial intelligence observation station network layout optimization method combining PID and feature contribution analysis. It is applied to the field of station network layout technology, including data acquisition and preprocessing, multi-factor representation and target function construction, swarm intelligence optimization, feature contribution analysis, and layout scheme evaluation and output. In this way, the present application integrates multi-source heterogeneous data, uses the PID algorithm of secondary interpolation optimization to make station network optimization decisions, and quantifies the contribution of each factor by combining the SHAP explainability model. It can significantly improve the scientificity, intelligence and decision transparency of the layout, avoid local optimization, and achieve a comprehensive balance between economic benefits and monitoring benefits.

Description

Technical Field

[0001] This disclosure relates to the field of station network deployment technology, and in particular to an artificial intelligence observation station network layout optimization method that integrates PID and feature contribution analysis. Background Technology

[0002] With the development of meteorological monitoring technology, optimizing the layout of meteorological stations has become an important means to improve the accuracy and coverage of meteorological data collection. However, existing station network deployment methods still have shortcomings in terms of intelligence, multi-factor trade-off analysis, and model interpretability, which limits their application effectiveness in complex environments.

[0003] The aforementioned problems indicate that existing station network deployment technologies still have significant shortcomings in terms of intelligent optimization, multi-factor trade-off analysis, and model interpretability. How to construct an optimization model that can intelligently and efficiently search for the optimal solution and clearly reveal the contribution of each decision factor has become a critical challenge that urgently needs to be addressed. Summary of the Invention

[0004] This disclosure provides an AI-based method for optimizing the layout of an observation network that integrates PID and feature contribution analysis. By integrating multi-source heterogeneous data, applying a swarm intelligence optimization method based on PID algorithm and quadratic interpolation optimization, and combining it with the SHAP interpretability model, the method achieves the globally optimal layout of the network and quantifies the contribution of each factor to the layout decision. This solves the technical challenges of existing network layout technologies in terms of intelligence level, multi-factor trade-off analysis, and model interpretability.

[0005] According to a first aspect of this disclosure, an artificial intelligence observation network layout optimization method integrating PID and feature contribution analysis is provided, comprising the following steps; Acquire various types of raw data required for the deployment of the meteorological observation station network, and clean, integrate, verify, validate, and standardize the raw data; The preprocessed multivariate data is transformed into quantifiable feature vectors, and a fitness function containing objective function and constraints is constructed according to business needs and optimization objectives. Based on the fitness function, an improved PID search optimization algorithm is used to obtain the globally optimal or near-optimal station network layout scheme. An interpretability analysis was performed on the optimized station network layout scheme to quantify the contribution of each influencing factor to the final evaluation result and obtain the optimal layout scheme. The optimal deployment scheme is then evaluated to obtain the evaluation results.

[0006] Preferably, the raw data includes: geospatial data, meteorological data, socioeconomic data, and existing station network operation data; The geospatial data includes high-precision digital elevation model data, land use / cover data, and administrative division boundary data. The geospatial data is used to extract topographic features such as altitude, slope, topographic relief, and water system distribution within the region, and to identify the land cover. The meteorological data includes historical wind profile data, historical severe weather event data, water vapor channel data, and real-time meteorological data. The meteorological data is used to obtain the distribution information of elements such as wind speed, wind direction, temperature, and humidity in the region, the spatial distribution and impact intensity of severe weather events, water vapor transport paths, and real-time meteorological elements. The socioeconomic data includes population density distribution data, GDP distribution data, and infrastructure distribution data. This socioeconomic data is used to assess the coverage capacity of the meteorological network in densely populated areas, the demand for meteorological services from regional economic activities, and the convenience and cost of the construction and maintenance of the meteorological network. The existing station network operation data includes operational parameters such as the location coordinates of existing wind stations, detection range, historical data integrity, and equipment maintenance costs.

[0007] Preferably, the process of acquiring various types of raw data required for the deployment of the meteorological observation station network, and cleaning, integrating, verifying, and standardizing the raw data, is as follows: Establish connections with various data sources and obtain raw data through standard APIs or data exchange protocols; The collected raw data is cleaned by performing redundant data identification and deletion, missing value filling, outlier detection and correction, and data format standardization. The cleaned multi-source heterogeneous data is spatially registered and time-synchronized, and vector data and raster data of different formats are overlaid and fused through a geographic information system. The integrated data is numerically scaled to obtain preprocessed data.

[0008] Preferably, the feature vector includes: station location features, terrain features, meteorological features, socio-economic features, ecological environment features, existing station network coverage features, and station detection range features; The location features of the sites include the longitude, latitude, and altitude of each potential site; The terrain features include the average slope, topographic relief, and aspect of the area where the station is located. The meteorological characteristics include the historical average wind speed, wind direction frequency distribution, precipitation intensity, temperature variability, water vapor channel intensity index, and frequency and intensity index of severe weather events in the area where the station is located; The socioeconomic characteristics include the weighted average population density, weighted average GDP density, and the number or distance of important infrastructure within a certain radius around the site. The ecological environment characteristics include the vegetation coverage around the site and the ecological sensitivity zoning level. The existing station network coverage characteristics include the coverage blind area, overlapping coverage area, and monitoring data integrity of the existing station network; The site detection range characteristics include the effective detection radius of each site, which is affected by terrain, climate conditions, or equipment model.

[0009] Preferably, the fitness function includes the following constraints: coverage constraint, budget constraint, minimum distance constraint between sites, terrain and vegetation suitability constraint, and population coverage. The coverage constraint is used to ensure that all predefined target areas to be covered are covered by at least one wind station's effective detection range. Specifically, the study area is divided into M grid cells G. j Each site S i With a detection range D ( Si ), then for all grid cells G j There exists at least one site S i G j ⊆D(S i Alternatively, the constraint could be transformed into a threshold requirement for coverage. The budget constraint is used to ensure that the total construction cost, operating cost, and maintenance cost of all new wind stations do not exceed the preset total budget; The minimum distance constraint between sites is used to ensure that the geographical distance between any two different wind stations must be greater than a preset minimum threshold d. min ; The terrain and vegetation suitability constraints are used to ensure that wind stations cannot be located in specific unsuitable areas, including steep slopes with excessive gradients, high-risk areas for geological disasters, water source protection areas, ecological red line areas, areas with high-density vegetation cover, and areas where infrastructure support cannot be provided. The population coverage and service equity constraints are designed to ensure that areas with specific population densities are effectively covered and to achieve relative equity in meteorological services across areas with different population distributions.

[0010] Preferably, the fitness function is specifically: Wherein: F coverage (S) is the coverage benefit function, F cost (S) is a cost function, representing the total cost of the construction and operation of the station network, which is calculated based on the budget constraints.redundancy (S) is the redundancy function, Penalty k (S) represents the penalty term for the k-th constraint. When solution S violates a constraint, the corresponding penalty term introduces a large negative value to reduce the fitness of the solution. P is the number of constraints, and Fitness(S) is the fitness function, where α, β, γ, δ k These are the weighting coefficients for each objective and penalty.

[0011] Preferably, the process of obtaining the globally optimal or near-optimal station network deployment scheme using the improved PID search optimization algorithm based on the fitness function is as follows: Generate an initial population and set the population size. Each individual in the population represents a potential station network deployment scheme, which consists of the geographical locations of NN stations to be deployed. Each individual can be encoded as a 2N-dimensional vector, where the first N dimensions represent longitude and the last N dimensions represent latitude. For each individual in the population, its fitness value is calculated based on the fitness function, and the population is updated using an incremental PID controller method combined with a quadratic difference generation function. The process involves repeated iterations and updates. Once the termination conditions are met, the search stops, and a globally optimal station network layout scheme is generated. The termination conditions include: reaching the preset maximum number of iterations, the fitness value of the globally optimal solution no longer changing significantly in a number of consecutive iterations, or the fitness value of the found layout scheme reaching or exceeding the preset target accuracy requirement.

[0012] Preferably, the process of calculating the fitness value of each individual in the population based on the fitness function, and updating the population individuals using an incremental PID controller method combined with a quadratic difference generation function is as follows: In each iteration, the systematic bias is calculated based on the fitness function of the individual, where the systematic bias is the difference in fitness between the best individual in the current iteration population and the global best individual. Incremental PID output is calculated based on the system deviation. The calculated incremental PID output is used to update the parameters of each individual in the population; A secondary interpolation operation is triggered based on preset conditions, and the current optimal solution, the previous optimal solution, and the global optimal solution are selected as key data points. Using the key data points as function values ​​and a certain dimension of the key data points in the search space as independent variables, a quadratic function is generated and the extreme points are calculated to obtain new candidate solutions. The new candidate solution is introduced into the population as a new trial individual and iteratively updated with other individuals in the existing population.

[0013] Preferably, the interpretability analysis of the optimized station network layout scheme is implemented using the SHAP model, wherein: The target model of the SHAP model is a fitness function or its surrogate model used to evaluate the station network scheme, the input feature is the quantifiable feature vector, and the prediction target is the comprehensive evaluation value of the station network deployment scheme. The SHAP model quantifies the importance of each feature by calculating the average marginal contribution of each feature to the model's prediction results across all possible feature combinations. For feature i, its Shapley value ϕ i The formula for calculating (f,x) is: Where K is the set of all features, |K| is the total number of features, S is the subset of features other than feature i, and f x (S) is the model's prediction when using only features from a subset S as input, calculated by setting features not in S as baseline values ​​in the target model.

[0014] Preferably, the final evaluation of the optimal deployment scheme adopts an evaluation index system including: coverage, data redundancy, construction and operation costs, data quality potential, geographical accessibility, and fitness value. The coverage rate represents the proportion of the target area that is effectively monitored; the data redundancy indicates whether there are stations with overlapping functions or highly related information in the station network; the construction and operation cost represents the overall construction investment and long-term operation and maintenance cost of the station network; the data quality potential represents the potential data quality obtained by evaluating the meteorological characteristics of the station location; and the fitness value is the final fitness value obtained by the optimization algorithm and is used as a comprehensive performance indicator.

[0015] Compared with the prior art, the present invention has the following technical effects: This invention integrates multi-source heterogeneous data, including geographic information, climate and environment, socio-economic data, land use data, disaster risk data, and existing infrastructure data, to construct a comprehensive and multi-dimensional feature set, significantly enhancing the representational capability of station network deployment decisions. Compared to existing technologies that rely solely on historical meteorological data or single geographic factors, this invention can more accurately reflect the combined impact of various factors in complex deployment environments, laying a solid data foundation for optimized decision-making.

[0016] This invention innovatively introduces a swarm intelligence optimization method based on the PID algorithm into the field of station network deployment, and integrates a quadratic interpolation optimization mechanism. The PID search algorithm, with its powerful global search capability and adaptive adjustment mechanism for system deviations, can efficiently find the globally optimal station network deployment scheme in a multi-dimensional, discrete, and complexly constrained search space. The quadratic interpolation optimization mechanism effectively enhances the algorithm's local fine-grained optimization capability, significantly avoiding the defect of traditional swarm intelligence algorithms easily getting trapped in local optima, and ensuring the quality and accuracy of the deployment scheme. This overcomes the limitations of existing technologies that rely on empirical judgment or local optimization leading to suboptimal deployment schemes.

[0017] This invention introduces the SHAP interpretability model to deeply analyze the optimization decision-making process. The SHAP model can quantify and visualize the contribution of multiple factors such as terrain, climate, population, and economy to the final station network deployment decision. This transparent and interpretable decision-making mechanism solves the problems of insufficient interpretability and "black box" operation in existing technical models, enhances decision-makers' trust in the deployment plan, and provides a strong scientific basis for the adjustment, demonstration, and promotion of the plan. By clarifying the contribution ratio of each factor, decision-makers can gain a deeper understanding of "why the deployment is done this way," thereby making more informed and targeted plans.

[0018] The fitness function constructed in this invention comprehensively considers multiple interdependent objectives, such as maximizing coverage, minimizing construction and operation costs, maximizing information gain, and minimizing disaster risk exposure. It also incorporates operational constraints, such as coverage constraints, budget constraints, and minimum distance constraints between sites, into the fitness function as penalty terms. This multi-objective, multi-constraint optimization framework ensures a comprehensive balance between economic benefits, monitoring benefits, environmental adaptability, and security in the final deployment scheme, significantly improving the overall performance and resource utilization efficiency of the network and avoiding the risks of data redundancy and data blind spots.

[0019] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A schematic flowchart of an artificial intelligence observation network layout optimization method integrating PID and feature contribution analysis according to an embodiment of the present disclosure is shown. Figure 2 This diagram illustrates a data source type of an artificial intelligence observation network layout optimization method that integrates PID and feature contribution analysis according to an embodiment of the present disclosure; Figure 3 The diagram illustrates the core execution steps of a PID algorithm for optimizing the layout of an artificial intelligence observation station network, which integrates PID and feature contribution analysis, according to an embodiment of this disclosure. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] This invention provides an artificial intelligence-based method for optimizing the layout of an observation network that integrates PID and feature contribution analysis, comprising the following steps: S101. Obtain various types of raw data required for the deployment of the meteorological observation station network, and clean, integrate, verify, validate and standardize the raw data.

[0024] Regarding data types and sources, this embodiment collects five major categories of data: geospatial data, meteorological data, socioeconomic data, ecological and environmental data, and existing station network operation data, specifically as follows: Figure 2As shown. Specifically, geospatial data includes Digital Elevation Model (DEM) data with a spatial resolution of 30 meters, typically sourced from the National Geographic Information Public Service Platform or SRTM Global Digital Elevation Data. This data is used not only to extract topographic factors such as elevation, slope, aspect, and topographic relief, but also, in conjunction with water system distribution data, to identify potential runoff paths and catchment areas. Land use / cover (LULC) data, with a spatial resolution of 10 meters, is typically interpreted using European Space Agency (ESA) Sentinel-2 or USGS Landsat satellite imagery products. This data is used to accurately identify various land cover types, including forests, grasslands, farmland, water bodies, and urban construction land. Administrative boundary data, usually sourced from standard datasets published by national civil affairs or natural resources departments, is used to define the study area and assist in the calculation and management of regional attributes. Meteorological data includes historical wind profile data, sourced from observation equipment such as upper-air sounding stations, wind profiler radars, and microwave radiometers. Its temporal resolution is typically set to 1 hour, with a resolution finer than 100 meters, providing distribution information on elements such as wind speed, wind direction, temperature, and humidity, which is crucial for station network planning. Historical data on severe weather events, such as the spatial distribution, frequency, and intensity of events like heavy rain, typhoons, and severe convection, can be obtained from the National Meteorological Information Center database or historical disaster archives of regional meteorological departments. Water vapor channel data can be extracted from the output products of numerical weather prediction models (such as WRF and CMA-MESO models) or obtained through remote sensing inversion (e.g., MODIS or AIRS water vapor products), with a spatial resolution of up to 5 kilometers, used to characterize the potential paths and intensity of atmospheric water vapor transport within the region. Real-time meteorological data, including hourly surface wind speed, wind direction, temperature, and precipitation, comes from the regional automatic weather station network. In terms of socioeconomic data, population density distribution data, with a spatial resolution of up to 100 meters × 100 meters, is sourced from the National Bureau of Statistics' geographic information system census data or WorldPop global population distribution data. This data is used to assess the coverage capacity and service benefits of the meteorological network in densely populated areas. Gross Domestic Product (GDP) distribution data, with a spatial resolution typically of 1 kilometer × 1 kilometer, is obtained through the fusion of statistical yearbooks and spatial statistical interpolation products. This data is used to assess the demand for high-precision meteorological services from regional economic activities and the economic benefits of meteorological network construction. Infrastructure distribution data, such as road networks, power lines, and communication base stations, typically comes from transportation geographic information systems or infrastructure databases of land and resources departments. This data is used to assess the convenience and cost of meteorological network construction and maintenance. Ecological and environmental data includes vegetation index data (such as the Normalized Difference Vegetation Index (NDVI)). Its spatial resolution is similar to LULC data, typically 30 meters, and it is sourced from Landsat satellite remote sensing imagery. This data is used to assess vegetation cover and growth vitality.Ecological sensitivity zoning data, derived from zoning schemes released by the ecological and environmental departments, is used to avoid ecologically fragile or protected areas during the deployment process. Finally, existing station network operation data includes key operational parameters such as the location coordinates of existing wind stations, effective detection range, historical data integrity statistics, equipment maintenance costs, and operational failure rates.

[0025] Specifically, in the data cleaning and integration stage, this embodiment identifies and addresses common issues in the collected data, such as missing values, outliers, duplicate data, and inconsistent formats. For missing values, for numerical data, cubic spline interpolation or a time-series-based ARIMA model is used to imputation, fully utilizing the spatiotemporal correlation of the data; for categorical data, mode imputation or prediction imputation based on decision trees and other models is used. For outliers, machine learning algorithms based on isolation forests or local outlier factors (LOF) are used for identification, and deletion, correction (e.g., replacing with the mean or median of neighboring points), or labeling is selected based on the degree of deviation and business impact. Data deduplication is achieved by constructing a unique identifier based on a combination of site ID, observation time, and geographic coordinates to accurately delete duplicate records. Unified data format is key to ensuring system interoperability. All geospatial data is forcibly converted to GeoJSON or Esri Shapefile format, time series data is uniformly adopted in CSV or Parquet format, and all timestamps are specified to be Coordinated Universal Time (UTC). The geographic coordinate system is WGS84.

[0026] In the data verification and validation phase, to ensure data accuracy and logical consistency, this embodiment employs cross-validation. This involves comparing the consistency of data from different sources within the same geographical area. For example, it compares the station elevations extracted from DEM data with the measured elevation data from ground meteorological stations, marking stations with differences exceeding ±5 meters for further verification. Logical consistency checks ensure that the data conforms to physical laws and business logic. For instance, wind speed should be within a reasonable range (usually below 5 m / s), precipitation should not be negative, and relative humidity should not exceed 100%. Statistical testing involves performing descriptive statistical analysis on key data fields (such as mean, standard deviation, skewness, and kurtosis) and comparing them with historical statistical distributions. If the distribution characteristics significantly deviate from expectations, a data quality warning is triggered.

[0027] Finally, in the data normalization and encoding stage, to eliminate the influence of data with different dimensions and scales on model training and to adapt to the algorithm's requirements for processing discrete inputs, this embodiment uses Z-score normalization for numerical data normalization. This converts the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, with the formula x′=(x-μ) / σ. This method is relatively insensitive to outliers and better preserves the distribution characteristics of the original data. For certain specific scenarios, Min-Max normalization can also be chosen to scale the data to the [0, 1] interval, with the formula x′=(xx... min ) / (x max -x min In terms of categorical data encoding, for unordered discrete categorical features (such as land use type, different administrative divisions), a One-Hot encoder is used to convert them into binary vectors, effectively avoiding the model's misunderstanding of the "distance" between categories; for ordered categorical features (such as ecological sensitivity level), label encoding is used to convert them into numerical values, but it is necessary to ensure that the numerical order matches the category order.

[0028] S102. Transform the preprocessed multivariate data into quantifiable feature vectors, and construct a fitness function containing objective function and constraints based on business needs and optimization objectives.

[0029] This step transforms the preprocessed multivariate data into quantifiable feature vectors and constructs a fitness function based on preset business requirements and optimization objectives. This fitness function includes a clear objective function and a series of constraints to guide the optimization process of the swarm intelligence optimization algorithm.

[0030] In one specific embodiment, the factors related to the station network planning and layout are quantified and characterized as a series of computable features. Station location features are directly encoded as the longitude, latitude, and altitude extracted from DEM data for each potential station. Topographic features are calculated a second time based on high-resolution DEM data, including the average slope of the area where the station is located, calculated by averaging the elevation change rate within a 3×3 or 5×5 window, expressed in degrees or percentages; the Terrain Ruggedness Index (TRI), measured by the local elevation standard deviation or coefficient of variation, reflecting the ruggedness of the surface; and the Aspect, divided into 0-360 degrees, or discrete into eight directions, used to assess the station's exposure to or shading from specific wind directions. The extraction of meteorological characteristics is more complex. For example, the historical average wind speed of the station's area is obtained through spatial interpolation and statistical averaging of high-resolution wind field data over many years and seasons, typically including wind speeds at different heights such as 2 meters, 10 meters, and 70 meters; the wind direction frequency distribution is obtained by statistically analyzing the frequency of wind direction data in 16 directions over many years; precipitation intensity and temperature variability are characterized by monthly and seasonal averages; the water vapor transport index is obtained by integrating and normalizing water vapor flux data; and the hazardous event frequency and intensity index is calculated by overlaying historical disaster data with geographical location and using methods such as kernel density estimation. In terms of socioeconomic characteristics, this embodiment obtains the weighted average population density and the weighted average GDP density within a certain radius (e.g., 5 km, 10 km, 20 km) around the site, where the weights typically decrease with distance to reflect the impact on the service benefits of the surrounding area; and quantifies the number of important infrastructures (such as airports, large industrial parks, ports) or the distance to the nearest infrastructure to assess the economic benefits and service targets of the site deployment.Ecological and environmental characteristics include the vegetation coverage rate around the site, which is usually measured by the average of the NDVI or leaf area index (LAI); and the ecological sensitivity level, a discrete value that reflects the degree of ecological vulnerability of the area.

[0031] The existing station network coverage characteristics include evaluation indicators such as the blind spot area, redundant coverage area, data completeness, and data timeliness. These indicators are quantified using spatial analysis tools (such as Voronoi diagrams and buffer analysis) and historical data records. The station detection range characteristics define the effective detection radius of each wind station. This radius is a dynamic parameter, limited by equipment model, terrain, climate conditions (such as atmospheric visibility and turbulence intensity), and radio transmission loss, and needs to be determined through field calibration or electromagnetic wave propagation simulation models.

[0032] Furthermore, to quantify the impact of different factors on site layout, this embodiment assigns an initial weight parameter to each quantified factor. These weight parameters are not fixed but are optimized and adjusted through two mechanisms. First, the initial weight settings can be determined by the experience of domain experts (including meteorologists, geographic information system experts, regional economic planners, and ecological environmental protection experts), and quantified using the Analytic Hierarchy Process (AHP) to construct a judgment matrix and calculate the relative weights of each factor. For example, in a certain region, the weight of "average wind speed" might be set to 0.25, "population density" to 0.20, "construction cost" to 0.15, and "topographic relief" to 0.10, etc. Second, these weight parameters can be used as part of the optimization variables during the iteration process of the swarm intelligence optimization algorithm, and are adaptively adjusted through the algorithm's feedback mechanism. Specifically, when the algorithm finds that a certain weight combination leads to a higher fitness value, these weights will be strengthened; conversely, they will be weakened. This mechanism means that while searching for the optimal site layout, the algorithm is also learning and optimizing the relative importance of different factors, thereby significantly enhancing the model's adaptability and generalization ability to complex and changing environments.

[0033] In a preferred embodiment of the present invention, constraint design is key to modeling multi-objective optimization problems. This embodiment models the station network deployment problem as a multi-constraint optimization problem, specifically including: (1) Coverage constraint (C) covThis ensures that all predefined target areas are covered by the effective detection range of at least one wind station. The study area is first divided into M grid cells G, each with a side length of 1 km × 1 km. j Each grid cell represents a potential target area to be covered. Let N be the number of wind stations to be deployed, and S be the number of stations. i It has an effective detection radius R determined by the equipment model, terrain, and weather conditions. i Its detection range is defined as S i Centered on R i A circular region D(S) with radius S i Then the constraint is expressed as: ∀j∈{1,…,M},∃i∈{1,…,N} such that G j ∩D(S i )≠∅. In practical calculations, this constraint translates into a threshold requirement for coverage; for example, at least 95% of the grid cells in the total area to be covered must be covered, i.e.: (1) (2) Budget constraint (C budget The total construction, operation, and maintenance costs of all newly added wind stations must not exceed the pre-set total budget. Let c... build,i For site S i Construction costs (including equipment procurement, civil construction, power supply and communication facilities, etc.), c op,i For site S i The annual operating and maintenance costs (including energy consumption, personnel, inspections, etc.) are given, where T is the planning period (e.g., 10 years). The constraint is then expressed as: (2) Among them, B total The total budget is in RMB. Typically, construction costs range from RMB 2 million to RMB 5 million, and annual operating costs range from RMB 100,000 to RMB 300,000, depending on the equipment model, geographical location, and infrastructure conditions.

[0034] (3) Minimum distance constraint between stations (C) dist The geographical distance between any two different wind stations must be greater than a preset minimum threshold d. min This is to avoid data redundancy and resource waste, and to reduce electromagnetic interference between devices. For any two different sites S i and S k Its geographical distance dist(S i ,S k (The spherical distance calculated using the Haversine formula) must satisfy dist(S) i ,S k)≥d min Typically, d min The range of the detection equipment and the characteristics of the research area are set according to the effective range of the detection equipment. For example, it can be set to 30 kilometers in plains areas, while it may be appropriately reduced to 15-20 kilometers in complex mountainous areas.

[0035] (4) Topographic and vegetation suitability constraints (C geo Wind stations cannot be located in specific unsuitable areas. These unsuitable areas include: steep slopes with excessive gradients (e.g., slopes greater than 30 degrees), which may lead to high construction difficulty and maintenance risks; high-risk areas for geological hazards (e.g., areas prone to landslides and debris flows), identified through geological hazard risk layers; areas where construction is explicitly prohibited by laws and regulations, such as water source protection areas, ecological red line areas, and nature reserves; areas with high-density vegetation cover (e.g., forests or wetlands, where vegetation height may interfere with radar detection signals), such as areas with vegetation height exceeding 5 meters or canopy density exceeding 80%; and remote areas where basic infrastructure support (such as power supply and communication networks) cannot be provided. This constraint is achieved by constructing a series of geospatial Boolean masks, meaning that if a candidate point is located within any of the prohibited areas, its suitability score is 0.

[0036] (5) Population coverage and service equity constraints (C pop This ensures effective coverage of areas with specific population densities and achieves relative fairness in meteorological services across different population distribution areas. For example, all densely populated areas are divided into three levels based on population density: high-density areas (>5000 people / km²), medium-density areas (1000-5000 people / km²), and low-density areas (<1000 people / km²), with coverage rates of 98%, 95%, and 90% required for each level, respectively.

[0037] The fitness function is a crucial criterion for evaluating the performance of individual site layout schemes in optimization algorithms. This embodiment transforms the multi-objective optimization problem into a single-objective fitness function, integrating multiple objective functions and penalty terms through weighted summation to achieve comprehensive optimization. (3) Where: S represents a specific network deployment scheme, i.e., a set of N geographical locations {(Lon1,Lat1),…,(Lon N ,Lat N )}.

[0038] F coverage(S) is the coverage benefit function, used to quantify the effective coverage of a network of stations over a target area. This function is obtained by calculating the ratio of the number of grid cells covered by at least one station's detection range to the total number of grid cells to be covered. Different weights can be assigned based on the importance of different areas (such as population density and economic contribution) to calculate the weighted coverage rate. For example, its value is a floating-point number between [0, 1], representing the coverage rate.

[0039] F cost (S) is the cost function, representing the total cost of station network construction and operation. Its calculation method is as described in the budget constraint above. It is usually normalized to match the dimensions of other terms in the fitness function, for example, by converting it to a cost ratio in the range of [0, 1].

[0040] F redundancy (S) is the redundancy function, used to quantify the degree of overlapping coverage of a network of stations or the correlation between data between stations. It can be measured by calculating the ratio of the sum of the overlapping areas of the detection ranges of any two stations to the total coverage area, or by evaluating the correlation between data between stations through statistical methods (such as spatial autocorrelation analysis) to maximize information gain.

[0041] Penalty k (S) represents the penalty term for the k-th constraint. When solution S violates a constraint, the corresponding penalty term introduces a large negative value to significantly reduce the fitness of the solution. For example, if the budget constraint is violated, Penalty... budget It can be defined as the product of the square of the overspending amount and a larger coefficient; if the minimum distance constraint is violated, Penalty dist This can be defined as the product of the number of station pairs violating the distance threshold and a large coefficient. For terrain and vegetation suitability constraints, if a station is located in a prohibited area, the penalty term is directly taken as a very large negative value (e.g., -10). 6 This essentially rendered the plan obsolete.

[0042] α,β,γ,δ k Weighting coefficients are assigned to each objective and penalty, and these coefficients are adjusted based on expert experience (such as the Analytic Hierarchy Process) and practical needs to balance the priorities among different objectives. For example, in a scenario where coverage is the primary objective, the weight of α may be much higher than that of β and γ. In a scenario with an extremely tight budget, the weight of β will be increased accordingly.

[0043] By maximizing this fitness function, the optimization algorithm in this embodiment can effectively select station network deployment schemes that have wide coverage, low cost, low redundancy, and strictly meet all constraints.

[0044] S103. Based on the fitness function, use the improved PID search optimization algorithm to obtain the globally optimal or near-optimal station network layout scheme.

[0045] This step, based on multi-factor characterization and fitness functions, employs an improved PID-based Search Algorithm (PSA) to search for and discover the globally optimal or near-optimal station network deployment scheme. Specifically, as follows... Figure 3 As shown.

[0046] Specifically, regarding the population initialization mechanism, upon algorithm startup, an initial population is first generated, with a size set to PopSize (e.g., 200 to 500 individuals). Each individual represents a potential station network deployment scheme, consisting of the geographical locations (longitude Lon, latitude Lat) of N stations to be deployed. Each individual can be encoded as a 2N-dimensional vector, where the first N dimensions represent longitude and the last N dimensions represent latitude. For each station to be deployed, its longitude and latitude are uniformly and randomly initialized as two independent parameters within the predefined study area boundary. For example, the longitude range is [Lon...]. min Lon max ], latitude range is [Lat min ,Lat max The strict setting of parameter boundaries ensures that the initial site locations fall within a reasonable and deployable geographical range; that is, all sites must be located within the administrative boundaries of the study area. Random initialization ensures that the initial population has sufficient diversity to cover a broad search space, thus effectively preventing the algorithm from prematurely getting trapped in local optima.

[0047] Furthermore, in the fitness function calculation stage, for each individual in the population (i.e., each deployment scheme), this embodiment calculates its fitness value based on the fitness function. The fitness value reflects the quality of the scheme; a higher fitness value indicates a better deployment scheme. The calculation process strictly adheres to all constraints, and penalty terms are used to handle constraint violations, ensuring that the algorithm always searches towards the optimal solution within the feasible solution space.

[0048] Furthermore, this embodiment introduces an incremental PID (Proportional-Integral-Derivative) controller method to drive the updating of individuals in the population. The PID controller generates a control output by weighting and combining the error (system deviation), the integral of the error, and the derivative of the error, in order to adjust the system state.

[0049] Specifically, the system error calculation is crucial: in each iteration t, the calculation of the error e(t) is the core of PID control. In this embodiment, the error e(t) is defined as the best individual X in the population during the current iteration. current_best With the best individual X in the world global_best Differences in fitness between them: (4) This bias quantifies the gap between the current population and the historical best solution, with the goal of bringing the current population closer to the global best solution.

[0050] Incremental PID output (Δu(t)) calculation: The output of an incremental PID controller is the increment of the control quantity, not the absolute control quantity. The formula is: (5) Among them, K p K is the proportionality coefficient. i K is the integral coefficient. d These are the differential coefficients. These coefficients are pre-defined hyperparameters, such as K. p ∈[0.1,1.0], K i ∈[0.01,0.1], K d K ∈[0.05,0.5], and its specific value will be optimized through experiments. p The term reflects the immediate impact of the current bias on the update amount, used to accelerate convergence; K i The term reflects the cumulative effect of historical bias and is used to eliminate steady-state error and enhance the globality of the search; K d This term reflects the trend of deviation changes and is used to suppress oscillations and improve the stability of the algorithm.

[0051] Individual parameter update: The calculated Δu(t) is used to update the parameters of each individual X in the population. i The parameters of (t). Each individual X i The parameter update rule for (t+1) can be expressed as: (6) Among them, X i (t) represents the parameter vector of individual i at iteration t (i.e., the latitude and longitude of all stations), X global_best This is the parameter vector of the currently discovered global optimum, where λ is a learning rate or step size factor used to control the update magnitude. Represents individual X i (t) is the direction vector pointing to the global optimal solution. It is a term that introduces random perturbation, where R randIt is a random vector that follows a standard normal distribution. σ is a parameter that controls the amplitude of the perturbation, used to enhance the algorithm's exploration ability and avoid premature convergence. The updated individual parameters (longitude, latitude) will be strictly limited to their upper and lower bounds. If they exceed the boundaries, they will be pruned or mapped back to the valid interval. For example, if the maximum longitude is exceeded, the maximum longitude value will be taken.

[0052] To effectively prevent the PID search algorithm from getting trapped in local optima, this embodiment incorporates a quadratic interpolation function during the algorithm iteration process. This enhances the algorithm's ability to refine local searches and escape local optima. The quadratic interpolation operation is not performed at every step, but is triggered based on preset conditions. Specifically, the triggering mechanism can be: periodic triggering, for example, performing interpolation every 50 iterations; or stagnation detection triggering, when the fitness value of the global optimum changes less than a preset threshold (e.g., 10) over 20 consecutive iterations. -5 When the optimal solution (X) is reached, interpolation is triggered. The selection of interpolation data points is the core of quadratic interpolation. In this invention, the three key data points selected are: the current optimal solution (X...). current_best ), which is the best individual in the population in the current iteration; the best solution in the previous search (X previous_best ), that is, the best individual at the time the interpolation operation was last triggered; and the global optimal solution (X). global_best The fitness values ​​of these three solutions are the best individuals from all iterations so far. The fitness values ​​of these three solutions will be used as function values, while their dimensions in the search space (e.g., representative coordinates after dimensionality reduction, such as principal components of PCA) will be used as independent variables. A quadratic function f(x) = ax will be fitted using these three sets of points (e.g., (x1,f1),(x2,f2),(x3,f3)). 2 +bx+c. Then calculate the extreme points (vertices) x of this quadratic function. new =-b / (2a). This extreme point represents the potential optimal solution obtained by parabolic approximation within the current local region. The new candidate solution X generated by quadratic interpolation... new (All dimensions have undergone similar processing.) It is introduced into the population as a new, exploratory individual. This integration can involve directly replacing the worst individual in the population, adding it as an additional individual to expand the population size, or even crossbreeding it with the current best solution to generate a new individual with interpolation advantages. Through this mechanism, quadratic interpolation can guide the algorithm to perform a more refined search in local regions and use the discovered optimal information to infer more promising solutions, thereby effectively helping the algorithm escape local optima and improving its global optimization ability.

[0053] Finally, the PID search algorithm iterates through the above steps until a preset termination condition is met. The termination condition typically includes: the algorithm reaching a preset maximum number of iterations (e.g., 5000); and the fitness value of the global optimum no longer changing significantly over several consecutive iterations (e.g., a change of less than 10 in 100 consecutive iterations). -6 This indicates that the algorithm has converged; or that the fitness value of the found deployment scheme has reached or exceeded the preset target accuracy requirement. Once either termination condition is met, the algorithm stops searching and returns the globally optimal station network deployment scheme found so far.

[0054] S104. Perform interpretability analysis on the optimized station network layout scheme, quantify the contribution of each influencing factor to the final evaluation result, and obtain the optimal layout scheme.

[0055] This step involves an interpretability analysis of the optimized station network deployment plan, quantifying the contribution of each influencing factor to the final evaluation result, thereby providing decision-makers with understandable quantitative evidence.

[0056] Specifically, in the SHAP model construction and integration phase, this embodiment uses the trained fitness function used to evaluate the network deployment scheme as the target model for the SHAP interpretation model. If the fitness function is too complex, making direct calculation of the SHAP value inefficient, a surrogate model can be constructed, such as Gradient Boosting Trees (XGBoost / LightGBM) or a Multi-layer Perceptron (MLP) neural network. This surrogate model can accurately simulate the behavior of the fitness function. In this embodiment, all quantified influencing factors (such as altitude, slope, population density, average wind speed, infrastructure distance, etc.) are defined as the input features of the SHAP model, totaling K features. The prediction target is the "prediction result" interpreted by the SHAP model, i.e., the comprehensive evaluation value (its fitness value) of a specific network deployment scheme. This embodiment deeply integrates the SHAP (Shapley Additive exPlanations) open-source library (implemented in Python), which provides efficient Shapley value calculation capabilities and supports multiple interpreters to adapt to different types of target models.

[0057] Furthermore, the Shapley value calculation principle and application are based on the Shapley value theory in cooperative game theory. It quantifies the importance of each feature by calculating the average marginal contribution of each feature to the model's prediction result across all possible feature combinations. For a given network layout scheme x and its corresponding feature vector x=(x1,x2,…,x…) K ), SHAP calculates each feature xi When considering the contribution of a feature, its marginal contribution across all possible feature subsets (consortium S) is taken into account. This means it considers the marginal contribution of a feature across all possible feature subsets (consortium S). i and without x i Under the given conditions, the model prediction results change. For feature i, its marginal contribution in the feature subset S is defined as f(S∪{i})-f(S), where f(S) is the model prediction value using only the feature subset S. The Shapley value ϕi(f,x) of feature i is calculated using the following formula: (7) Where K is the set of all features, |K| is the total number of features, S is the subset of features other than feature i, and f x (S) is the model's prediction using only features from a subset S as input. It is calculated by setting features not in S as baseline values ​​(such as the mean or zero) in the target model. In this system, f x (S) is the fitness function evaluation value of the station network scheme when only the features in S are considered. Since the number of features is large, exhaustively calculating the Shapley value by listing all subsets is computationally intensive (complexity O(K·2)). K This embodiment employs approximation algorithms such as KernelSHAP or TreeSHAP for efficient computation. KernelSHAP approximates the Shapley value by training a weighted linear regression model, with a computational complexity of approximately O(K⋅M⋅T⋅logT), where M is the number of samples and T is the depth of the tree. TreeSHAP is optimized for tree-based models (such as gradient boosting trees) and can be computed in polynomial time (e.g., O(K⋅T)). 2 )) Calculate the precise Shapley value.

[0058] As a crucial component of this invention, the feature contribution analysis and visualization, utilizing Shapley values ​​calculated through SHAP, provides multi-level interpretations for station network planning. In terms of global interpretation, by averaging or aggregating the Shapley values ​​of all deployment schemes or a representative dataset, the average contribution of each feature to the overall station network evaluation can be obtained. For example, it can be clearly shown that "topographic relief" contributes an average of 15% to the evaluation result, "population density" contributes 12%, and "average wind speed" contributes 20%, thus identifying the most important factors influencing station network deployment decisions. In terms of local interpretation, for each specific optimized deployment scheme (and even each station within the scheme), it can be clearly seen whether the contribution of each feature (e.g., the altitude of a specific station, surrounding wind speed, distance from roads, etc.) to its final fitness value (or evaluation result) is positive or negative, and the magnitude of the contribution. This local interpretation can answer specific questions such as "why this station location is so excellent (or poor)," for example, "the station has high fitness mainly because it is located in an area with high average wind speed and high population density, but its construction cost is slightly higher." Interaction effect analysis is another powerful feature of SHAP, which can reveal the interaction between features, that is, how the influence of one feature depends on the value of another feature. For example, through SHAP interaction values, it can be found that "in low-altitude areas, the contribution of wind speed to the effectiveness of the station network may not be as significant as in high-altitude areas, because the terrain at low altitudes is complex and greatly affected by building shading."

[0059] Finally, this embodiment intuitively presents decision-makers with the contribution ratio of each factor and the model's decision-making logic, and generates the optimal deployment scheme based on the analysis results of the contribution magnitude, effectively improving the model's transparency and credibility.

[0060] S105. Perform a final evaluation on the optimal deployment scheme and obtain the evaluation results. This embodiment employs a comprehensive evaluation index system to quantitatively evaluate its performance. These indexes include, but are not limited to: Coverage: The effective monitoring coverage area ratio of predefined key target areas (such as key meteorological sensitive areas, densely populated areas, economic development areas, and transportation hubs), which can be further refined into coverage rates for different area types.

[0061] Data redundancy: Does the network of stations have overlapping functions or highly related information? By using spatial clustering algorithms (such as DBSCAN) to identify overlapping areas or by using statistical methods such as Pearson correlation coefficient and mutual information to assess the correlation between data between stations, the amount of duplicate information can be quantified.

[0062] Construction and operation costs: The total investment in the construction of the station network (including land, equipment, and construction) and the long-term operation and maintenance costs (including energy consumption, personnel, spare parts, and periodic calibration) are given specific figures through a detailed cost accounting model.

[0063] Data quality potential: Assess the potential data quality and integrity based on the meteorological characteristics of the site location (such as average wind speed stability, turbulence characteristics, and the impact of topography on the wind field), for example, the average value of the predicted wind speed measurement error.

[0064] Geographic accessibility: Whether the site location is convenient for construction and subsequent maintenance, quantified by calculating the distance from the site to the nearest road or town, assessing terrain complexity and infrastructure conditions.

[0065] Fitness value: The final fitness value obtained by the optimization algorithm serves as a comprehensive performance indicator, directly reflecting the overall quality of the solution.

[0066] Furthermore, this embodiment demonstrates the advantages of the optimized solution in improving coverage, reducing costs, and improving data quality by quantitatively comparing the two in the aforementioned evaluation indicators. For example, the optimized solution improves coverage by X%, reduces costs by Y%, and reduces data redundancy by Z% compared to the existing network. This verifies the effectiveness and superiority of the present invention and provides strong support for decision-makers.

[0067] In a preferred embodiment of the present invention, the final deployment plan will be output in multiple forms to meet the needs of different decision-makers: Geographic Information System (GIS) Visualization: Optimized site locations are displayed in layers on a GIS platform (such as ArcGIS, QGIS, or a customized WebGIS platform), overlaid with background layers such as terrain, population distribution, land use, and water systems, intuitively presenting the spatial distribution and surrounding environment of the sites. Detailed attribute information for each site can be viewed by clicking on it.

[0068] Detailed parameter report: Generate a report containing detailed coordinates (longitude, latitude, altitude) of each station, estimated construction cost, estimated annual operating cost, a list of covered target areas, and the types of key meteorological phenomena to be monitored (such as low-level jet stream, boundary layer turbulence), and output it in PDF or Excel format.

[0069] Shapley Value Contribution Report: Provides a report based on SHAP analysis that quantifies the contribution of various influencing factors (such as terrain, wind speed, and population density) to the final optimization decision, and provides a local explanation of the site selection rationale for a specific site. For example, "Site A was selected here mainly because its high average wind speed contributed 0.15 to the fitness improvement, while its slightly higher construction cost deducted 0.05." Interactive Decision-Making Interface: Provides a user-friendly web or desktop interactive interface that allows decision-makers to adjust certain parameters or weights (e.g., budget constraints, coverage priorities) in real time, and instantly view changes to the proposed solution and their impact on the evaluation results and Shapley score. This interface can also perform sensitivity analysis, showing the differences in optimal solutions under different parameter settings, thereby assisting in the final decision.

[0070] Example This embodiment aims to illustrate in detail the practical application and effect of an artificial intelligence observation network layout optimization method that integrates PID and feature contribution analysis in a specific region, so as to verify the technical advancement and practicality of the present invention.

[0071] Background setting: Suppose this system is applied to a province in China (e.g., Province Z in East China). This province has complex terrain, including plains, hills, and coastal areas, with uneven population distribution and significant differences in economic development levels. It also faces threats from various meteorological disasters such as typhoons and rainstorms. Province Z currently has 15 wind stations, but their distribution is uneven, resulting in significant monitoring blind spots and data redundancy. Furthermore, many of these stations were built a long time ago, and some equipment is outdated. Therefore, there is an urgent need for network optimization and upgrades, and plans are in place to add 10 more wind stations.

[0072] Implementation steps: 1. Data Acquisition and Preprocessing: Geospatial data: 30-meter resolution DEM data of Z province (from the National Geomatics Center of China), 10-meter resolution LULC data (from ESA Sentinel-2), and administrative boundary data of Z province were collected.

[0073] Meteorological data: Hourly wind profile data for Z province over the past 10 years (100-meter resolution, covering altitudes from 100 meters to 3000 meters), data from automatic weather stations, 5-kilometer resolution water vapor flux data simulated by the WRF model, and historical typhoon tracks and rainfall event data.

[0074] Socioeconomic data: Collected population density data of 100m x 100m in Z province (sourced from WorldPop), GDP distribution data of 1km x 1km (sourced from the Z province statistics bureau and spatial statistics fusion), and distribution map of highways, railways and high-voltage transmission lines in Z province.

[0075] Ecological and environmental data: Obtain 30-meter resolution NDVI data, ecological red line and water source protection area vector data of Z province.

[0076] Existing station network operation data: Collect the location, detection range (set as a radius of 30 kilometers, but the actual range is affected by the terrain) of the 15 existing wind stations in Z province, historical data completeness rate (average of 90% over the past 5 years), and average annual maintenance cost (RMB 150,000 per station).

[0077] Data cleaning and integration: Missing values ​​in the wind profile data were processed using cubic spline interpolation. Anomalies in wind speed readings were detected and removed using the Isolation Forest algorithm. All data were standardized to the WGS84 coordinate system and UTC time.

[0078] Data normalization and coding: All numerical features (such as altitude, slope, wind speed, and population density) are Z-score standardized. LULC types and ecological sensitivity levels are One-Hot coded.

[0079] 2. Multi-factor characterization and objective function construction: Quantitative characteristics: A 1 km × 1 km grid was designated as candidate deployment points for Province Z. For each grid point, its altitude, average slope, TRI (Topographic Relief Index), average 10 m and 70 m wind speeds, dominant wind direction, frequency of heavy rainfall events in the past 10 years, weighted average population density within 10 km, weighted average GDP density, distance to the nearest highway, average NDVI, and whether it is located in an ecological red line area were calculated.

[0080] Factor weights: Initial weights were set using AHP combined with expert experience: average wind speed (0.25), coverage blind spot (0.20), construction cost (0.15), population density (0.10), terrain suitability (0.10), data redundancy (0.10), and others (0.10). These weights were adaptively adjusted by the algorithm during the optimization process.

[0081] Constraints: Coverage constraints (C cov ): Ensure that at least 98% of the area of ​​all regions in Province Z with a population density greater than 500 people per square kilometer and all major agricultural areas (classified as farmland by LULC) are covered.

[0082] Budget constraint (C budget The total construction and 10-year operation and maintenance cost of the 10 new sites will not exceed RMB 50 million. The average construction cost of each new site is RMB 2.5 million, and the annual operating cost is RMB 200,000.

[0083] Minimum distance constraint between stations (C) dist The distance between any two stations is not less than 25 kilometers.

[0084] Topographic and vegetation suitability constraints (Cgeo It is prohibited to set up stations in areas with a slope greater than 30 degrees, in mountainous areas with an altitude of more than 1,500 meters, in water source protection areas, and in areas with a forest coverage rate of more than 80%.

[0085] Population Coverage and Service Equity Constraints (C pop Ensure that the coverage rate of high-density areas (>5000 people / km²) and medium-density areas (1000-5000 people / km²) in Z Province reaches 99%, and the coverage rate of medium-density areas reaches 95%.

[0086] Fitness function: The previously defined formula is used. The penalty coefficient is set to 10 for budget overruns. -5 (Ten thousand yuan), the distance from the violation penalty coefficient is 10. 4 (Number), the penalty value for prohibited areas is -10. 6 .

[0087] 3. Swarm intelligence optimization based on PID algorithm: Population initialization: Set the population size PopSize=300. Encode the latitude and longitude of 10 new stations for each individual. The longitude range of the stations is [118°E, 123°E], and the latitude range is [27°N, 32°N]. The initial station locations are generated uniformly and randomly within the administrative region of Province Z.

[0088] PID parameters: After multiple experiments and optimizations, K was set to... p =0.8,K i =0.05,K d =0.2. Learning rate λ=0.01, random perturbation coefficient σ=0.001.

[0089] Quadratic interpolation fusion: Quadratic interpolation is triggered every 50 iterations. This occurs when the fitness value of the global optimum changes by less than 10 over 20 consecutive iterations. -5 Interpolation is also triggered at this time. The interpolated data points are the current best, the previous best, and the global best solutions, which are fitted along the longitude dimension to generate new candidate solutions and replace the worst individual in the population.

[0090] Iteration and Termination Conditions: The maximum number of iterations is set to 5000, or the fitness value of the global optimal solution changes by less than 10 over 100 consecutive iterations. -6 The process will end at that time.

[0091] 4. Feature contribution analysis based on the SHAP model: SHAP Model: Using the TreeSHAP interpreter, the target model is a surrogate model trained with LightGBM, which fits the fitness function with 99.5% accuracy.

[0092] Contribution Analysis: SHAP values ​​are calculated for the final optimized solution and typical solutions during the optimization process. The average contribution of each feature to the fitness value (global interpretation) and its contribution to specific site selection (local interpretation) are analyzed.

[0093] 5. Deployment plan evaluation and output: Evaluation metrics: Coverage of the optimized solution, data redundancy, total construction and operation costs, forecast data quality (measured by the average wind speed stability index), and average geographical accessibility (distance to the nearest road).

[0094] GIS Visualization: Display the precise locations (latitude and longitude) of 10 optimized new sites on a customized WebGIS platform, along with the existing 15 sites, overlaid with the DEM of Z Province and population density layers.

[0095] Detailed Report: Outputs a report containing detailed geographic coordinates of 25 sites, estimated construction and operation costs, main service areas, and forecast data quality.

[0096] SHAP Report: Submit a detailed SHAP contribution report, explaining that "average wind speed" contributed 22% to fitness improvement, "population density" contributed 18%, and "construction cost" deducted 10%, etc.

[0097] Final result of the optimization scheme (example) After 5,000 iterations (or early convergence), the system obtained the optimal deployment scheme for 10 newly added wind stations in Province Z.

[0098] Overall coverage: 99.2% coverage of densely populated areas and major agricultural areas.

[0099] Data redundancy: The overlap area of ​​the detection range of the stations in the network is 18.5%, which is significantly reduced compared with the traditional deployment scheme.

[0100] Total construction and 10-year operating costs: ultimately controlled within RMB 48.5 million, below the budget ceiling of RMB 50 million.

[0101] Average data quality potential: 85 points (out of 100, based on wind speed stability and turbulence index).

[0102] Geographical accessibility: The average distance from all new sites to the nearest road is 1.8 km, facilitating construction and maintenance.

[0103] Final fitness value: 0.875.

[0104] Example of SHAP feature contribution analysis: For the final optimized scheme, SHAP analysis showed that the feature that contributed the most to the fitness value was the "average 10-meter wind speed" (average Shapley value: +0.12), followed by the "population density weighted average" (+0.09), while "distance to the nearest highway" showed a slight negative contribution (-0.03), indicating that the site selection preferred areas with convenient transportation.

[0105] Comparative Example To further demonstrate the superiority of this invention, this comparative example will employ a station network deployment method based on a traditional genetic algorithm (GA) to plan 10 new wind stations in the same province Z region as the example, and compare it with the example. This traditional genetic algorithm does not include PID control concepts and quadratic interpolation fusion strategies; its individual updates mainly rely on selection, crossover, and mutation operations.

[0106] Traditional genetic algorithm deployment methods: 1. Data Acquisition and Preprocessing: The process is exactly the same as in the previous example, using the same multi-source data from Province Z and performing the same cleaning, normalization, and encoding processes.

[0107] 2. Multi-factor characterization and objective function construction: The process is exactly the same as in the example, using the same quantitative features, initial factor weights and constraints, as well as the same fitness function.

[0108] 3. Optimization of traditional genetic algorithms: Population initialization: Set the population size PopSize=300, encode the latitude and longitude of 10 new stations for each individual, and generate initial station locations uniformly and randomly within the administrative region of Z province, the same as in the previous example.

[0109] Fitness calculation: Similar to the example, the fitness value of each individual is calculated based on the defined fitness function.

[0110] Genetic manipulation: Selection: A roulette wheel selection method is used, where individuals with high fitness have a higher probability of being selected for the next generation.

[0111] Crossover: Single-point crossover is used, with a crossover probability set to 0.8. Two parent individuals are randomly selected and swapped at a random point in their parameter vectors (latitude and longitude list) to generate two child individuals.

[0112] Mutation: Uniform mutation is used, with a mutation probability set to 0.1. A parameter (longitude or latitude) of an individual is randomly selected and randomly replaced within a preset parameter range.

[0113] Iteration and Termination Conditions: The maximum number of iterations is set to 5000, or the fitness value of the global optimal solution changes by less than 10 over 100 consecutive iterations. -6 The process will end at that time.

[0114] 4. Deployment Scheme Evaluation and Output: The same evaluation index system as in the example is used to evaluate the optimal scheme obtained by the traditional genetic algorithm. SHAP feature contribution analysis is not performed because the traditional genetic algorithm lacks an embedded interpretability mechanism, and this comparative example mainly focuses on the quantitative comparison of optimization results.

[0115] The final result of the traditional genetic algorithm scheme (comparative example) After 5,000 iterations (or early convergence), the traditional genetic algorithm obtained the deployment scheme for 10 newly added wind stations in Province Z.

[0116] Overall coverage: 97.5% coverage of densely populated areas and major agricultural areas.

[0117] Data redundancy: The overlap area of ​​the detection range of stations in the station network accounts for 25.0%.

[0118] Total construction and 10-year operating costs: ultimately controlled at 52 million RMB, exceeding the budget limit.

[0119] Average data quality potential: rated at 80 points.

[0120] Geographical accessibility: The average distance from all new stations to the nearest road is 2.5 km.

[0121] Final fitness value: 0.810.

[0122] To visually demonstrate the superiority of the system provided by this invention, the following table quantifies and compares the key performance indicators of the embodiments and comparative examples, as shown in Table 1. Comparative analysis of results: As can be clearly seen from the comparison table above, the station network deployment system that integrates PID algorithm and feature contribution analysis provided by this invention demonstrates significant superiority in all key evaluation indicators.

[0123] First, regarding the overall coverage rate, the proposed solution achieves 99.2%, representing an absolute improvement of 1.7 percentage points compared to the 97.5% of the traditional genetic algorithm. This indicates that the proposed solution can more effectively cover the target area, especially in densely populated areas and key agricultural regions, ensuring more comprehensive meteorological service capabilities.

[0124] Secondly, data redundancy is an important indicator for measuring the efficiency of network resource utilization. The solution in this invention controls redundancy at 18.5%, a relative reduction of 26.0% compared to the comparative example of 25.0%. This means that the invention can plan sites more rationally, avoiding significant overlapping coverage and information waste, and improving the economic efficiency of network operation.

[0125] In terms of total cost control, the total construction and 10-year operating cost of this invention is RMB 48.5 million, successfully controlled within the budget limit of RMB 50 million, and achieving a relative cost reduction of 6.7% compared to the comparative cost of RMB 52 million. This directly demonstrates the invention's superior ability in resource optimization and significantly improves the economic feasibility of the project.

[0126] In terms of average data quality potential, the evaluation score of this invention is 85 points, higher than the comparative example's 80 points, representing a relative improvement of 6.25%. This indicates that when selecting a site, this invention tends to choose areas with stable wind speeds and suitable turbulence characteristics, thereby providing more accurate and higher-quality raw data for weather forecasting.

[0127] Regarding average geographical accessibility, the average distance of the proposed solution is 1.8 kilometers, significantly lower than the comparative example's 2.5 kilometers, representing a relative reduction of 28.0%. This not only reduces the difficulty of site construction but also greatly facilitates subsequent maintenance and repair, reducing operating costs and time investment.

[0128] The final fitness value, as a comprehensive performance indicator, shows that the proposed solution's 0.875 is significantly higher than the comparative example's 0.810, representing a relative improvement of 8.0%. This comprehensively reflects the overall advantages of the proposed solution in multi-objective trade-off optimization.

[0129] Furthermore, regarding the convergence speed of the algorithm, this invention employs a dynamic adjustment strategy of the PID algorithm and integrates quadratic interpolation, enabling the algorithm to converge in 3250 iterations, while the traditional genetic algorithm requires 4100 iterations. The convergence speed of the proposed solution is improved by approximately 20.7%. This demonstrates the significant advantage of this invention in optimization efficiency, enabling it to provide high-quality optimization solutions more quickly.

[0130] In summary, this invention significantly improves the global optimization capability and convergence speed of the optimization algorithm by introducing a dynamic adjustment mechanism and a quadratic interpolation strategy from the PID search algorithm, overcoming the tendency of traditional swarm intelligence algorithms to get trapped in local optima. Simultaneously, the comprehensive quantification of multiple factors and refined constraint design ensure the scientific validity and practicality of the solution. More importantly, the introduction of the SHAP model makes the complex optimization decision-making process transparent and interpretable, providing decision-makers with clear quantitative evidence. Therefore, it comprehensively surpasses traditional station network deployment methods in multiple dimensions, including intelligence, precision, transparency, economic benefits, and social benefits.

[0131] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0132] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0133] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for optimizing the layout of an artificial intelligence-based observation network that integrates PID and feature contribution analysis, characterized in that, Includes the following steps; Acquire various types of raw data required for the deployment of the meteorological observation station network, and clean, integrate, verify, validate, and standardize the raw data; The preprocessed multivariate data is transformed into quantifiable feature vectors, and a fitness function containing objective function and constraints is constructed according to business needs and optimization objectives. Based on the fitness function, an improved PID search optimization algorithm is used to obtain the globally optimal or near-optimal station network layout scheme. An interpretability analysis was performed on the optimized station network layout scheme to quantify the contribution of each influencing factor to the final evaluation result and obtain the optimal layout scheme. The optimal deployment scheme is then evaluated to obtain the evaluation results.

2. The method for optimizing the layout of an artificial intelligence observation station network according to claim 1, characterized in that, The raw data includes: geospatial data, meteorological data, socioeconomic data, and existing station network operation data; The geospatial data includes high-precision digital elevation model data, land use / cover data, and administrative division boundary data. The geospatial data is used to extract topographic features such as altitude, slope, topographic relief, and water system distribution within the region, and to identify the land cover. The meteorological data includes historical wind profile data, historical severe weather event data, water vapor channel data, and real-time meteorological data. The meteorological data is used to obtain the distribution information of elements such as wind speed, wind direction, temperature, and humidity in the region, the spatial distribution and impact intensity of severe weather events, water vapor transport paths, and real-time meteorological elements. The socioeconomic data includes population density distribution data, GDP distribution data, and infrastructure distribution data. This socioeconomic data is used to assess the coverage capacity of the meteorological network in densely populated areas, the demand for meteorological services from regional economic activities, and the convenience and cost of the construction and maintenance of the meteorological network. The existing station network operation data includes operational parameters such as the location coordinates of existing wind stations, detection range, historical data integrity, and equipment maintenance costs.

3. The method for optimizing the layout of an artificial intelligence observation station network according to claim 2, characterized in that, The process of acquiring various types of raw data required for the deployment of the meteorological observation station network, and cleaning, integrating, verifying, and standardizing the raw data, is as follows: Establish connections with various data sources and obtain raw data through standard APIs or data exchange protocols; The collected raw data is cleaned by performing redundant data identification and deletion, missing value filling, outlier detection and correction, and data format standardization. The cleaned multi-source heterogeneous data is spatially registered and time-synchronized, and vector data and raster data of different formats are overlaid and fused through a geographic information system. The integrated data is numerically scaled to obtain preprocessed data.

4. The method for optimizing the layout of an artificial intelligence observation station network according to claim 1, characterized in that, The feature vector includes: station location features, terrain features, meteorological features, socio-economic features, ecological environment features, existing station network coverage features, and station detection range features; The location features of the sites include the longitude, latitude, and altitude of each potential site; The terrain features include the average slope, topographic relief, and aspect of the area where the station is located. The meteorological characteristics include the historical average wind speed, wind direction frequency distribution, precipitation intensity, temperature variability, water vapor channel intensity index, and frequency and intensity index of severe weather events in the area where the station is located; The socioeconomic characteristics include the weighted average population density, weighted average GDP density, and the number or distance of important infrastructure within a certain radius around the site. The ecological environment characteristics include the vegetation coverage around the site and the ecological sensitivity zoning level. The existing station network coverage characteristics include the coverage blind area, overlapping coverage area, and monitoring data integrity of the existing station network; The site detection range characteristics include the effective detection radius of each site, which is affected by terrain, climate conditions, or equipment model.

5. The method for optimizing the layout of an artificial intelligence observation station network according to claim 4, characterized in that, The fitness function includes the following constraints: coverage constraint, budget constraint, minimum distance between sites constraint, terrain and vegetation suitability constraint, and population coverage. The coverage constraint is used to ensure that all predefined target areas to be covered are covered by at least one wind station's effective detection range. Specifically, the study area is divided into M grid cells G. j Each site S i With a detection range D ( Si ), then for all grid cells G j There exists at least one site S i G j ⊆D(S i Alternatively, the constraint could be transformed into a threshold requirement for coverage. The budget constraint is used to ensure that the total construction cost, operating cost, and maintenance cost of all new wind stations do not exceed the preset total budget; The minimum distance constraint between sites is used to ensure that the geographical distance between any two different wind stations must be greater than a preset minimum threshold d. min ; The terrain and vegetation suitability constraints are used to ensure that wind stations cannot be located in specific unsuitable areas, including steep slopes with excessive gradients, high-risk areas for geological disasters, water source protection areas, ecological red line areas, areas with high-density vegetation cover, and areas where infrastructure support cannot be provided. The population coverage and service equity constraints are designed to ensure that areas with specific population densities are effectively covered and to achieve relative equity in meteorological services across areas with different population distributions.

6. The method for optimizing the layout of an artificial intelligence observation station network according to claim 5, characterized in that, The fitness function is specifically: Wherein: F coverage (S) is the coverage benefit function, F cost (S) is a cost function, representing the total cost of the construction and operation of the station network, which is calculated based on the budget constraints. redundancy (S) is the redundancy function, Penalty k (S) represents the penalty term for the k-th constraint. When solution S violates a constraint, the corresponding penalty term introduces a large negative value to reduce the fitness of the solution. P is the number of constraints, and Fitness(S) is the fitness function, where α, β, γ, δ k These are the weighting coefficients for each objective and penalty.

7. The method for optimizing the layout of an artificial intelligence observation station network according to claim 1, characterized in that, The process of obtaining the globally optimal or near-optimal station network layout scheme using the improved PID search optimization algorithm based on the fitness function is as follows: Generate an initial population and set the population size. Each individual in the population represents a potential station network deployment scheme, which consists of the geographical locations of NN stations to be deployed. Each individual can be encoded as a 2N-dimensional vector, where the first N dimensions represent longitude and the last N dimensions represent latitude. For each individual in the population, its fitness value is calculated based on the fitness function, and the population is updated using an incremental PID controller method combined with a quadratic difference generation function. The process involves repeated iterations and updates. Once the termination conditions are met, the search stops, and a globally optimal station network layout scheme is generated. The termination conditions include: reaching the preset maximum number of iterations, the fitness value of the globally optimal solution no longer changing significantly in a number of consecutive iterations, or the fitness value of the found layout scheme reaching or exceeding the preset target accuracy requirement.

8. The method for optimizing the layout of an artificial intelligence observation station network according to claim 7, characterized in that, The process of calculating the fitness value of each individual in the population based on the fitness function, and updating the population individuals using an incremental PID controller method combined with a quadratic difference generation function is as follows: In each iteration, the systematic bias is calculated based on the fitness function of the individual, where the systematic bias is the difference in fitness between the best individual in the current iteration population and the global best individual. Incremental PID output is calculated based on the system deviation. The calculated incremental PID output is used to update the parameters of each individual in the population; A secondary interpolation operation is triggered based on preset conditions, and the current optimal solution, the previous optimal solution, and the global optimal solution are selected as key data points. Using the key data points as function values ​​and a certain dimension of the key data points in the search space as independent variables, a quadratic function is generated and the extreme points are calculated to obtain new candidate solutions. The new candidate solution is introduced into the population as a new trial individual and iteratively updated with other individuals in the existing population.

9. The method for optimizing the layout of an artificial intelligence observation station network according to claim 1, characterized in that, The interpretability analysis of the optimized station network layout scheme is performed using the SHAP model, wherein: The target model of the SHAP model is a fitness function or its surrogate model used to evaluate the station network scheme, the input feature is the quantifiable feature vector, and the prediction target is the comprehensive evaluation value of the station network deployment scheme. The SHAP model quantifies the importance of each feature by calculating the average marginal contribution of each feature to the model's prediction results across all possible feature combinations. For feature i, its Shapley value ϕ i The formula for calculating (f,x) is: Where K is the set of all features, |K| is the total number of features, S is the subset of features other than feature i, and f x (S) is the model's prediction when using only features from a subset S as input, calculated by setting features not in S as baseline values ​​in the target model.

10. The method for optimizing the layout of an artificial intelligence observation station network according to claim 1, characterized in that, The final evaluation of the optimal deployment scheme uses an evaluation index system that includes: coverage, data redundancy, construction and operation costs, data quality potential, geographical accessibility, and fitness value. The coverage rate represents the proportion of the target area that is effectively monitored; the data redundancy indicates whether there are stations with overlapping functions or highly related information in the station network; the construction and operation cost represents the overall construction investment and long-term operation and maintenance cost of the station network; the data quality potential represents the potential data quality obtained by evaluating the meteorological characteristics of the station location; and the fitness value is the final fitness value obtained by the optimization algorithm and is used as a comprehensive performance indicator.

Images