An artificial intelligence-based substation site selection method and system

Abstract

This invention discloses an artificial intelligence-based substation site selection method and system, relating to the field of substation site selection technology. The method includes: acquiring multi-source heterogeneous raw data containing remote sensing images, infrared thermal images, acoustic signatures, and text reports; constructing a multimodal site selection dataset through spatiotemporal benchmark alignment and preprocessing; employing a pre-trained cross-modal fusion neural network to extract features from each modality; filtering strongly correlated features through a cross-modal attention mechanism and fusing them to generate a globally correlated feature vector; finally, inputting this vector into a multi-task decision head; calculating a comprehensive score through weighted fusion; and outputting the site selection decision result. This invention solves the problem of inaccurate substation site selection methods in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of substation site selection technology, and in particular to a substation site selection method and system based on artificial intelligence. Background Technology

[0002] As the power system undergoes a profound transformation towards intelligence and digitalization, substations, as core hubs for power transmission and distribution, play an irreplaceable role in ensuring the reliability, economic efficiency, and regional energy allocation efficiency of the power system through their scientific and rational site selection. Substation site selection differs from that of ordinary commercial or logistics nodes; it requires not only consideration of spatial layout rationality but also strict adherence to multiple industry standards, including power planning, safety protection, and environmental adaptability. It is a complex systems engineering project integrating knowledge from multiple fields such as power engineering, geospatial science, environmental science, and policy regulations. Traditional substation site selection relies primarily on qualitative judgments based on expert experience, which is heavily influenced by subjective human factors, resulting in low efficiency, significant decision-making biases, and insufficient risk prediction, making it difficult to meet the demands of high-quality development in modern power systems. Against this backdrop, AI-driven site selection decision systems are gradually being applied to the field of substation site selection. Through data-driven intelligent analysis, they replace the traditional experience-driven model, becoming a core supporting technology for improving the accuracy and efficiency of substation site selection.

[0003] With the deep application of big data, cloud computing, artificial intelligence and Internet of Things technologies in the power sector, the data sources required for substation site selection are becoming increasingly abundant, forming a multi-source heterogeneous data system covering multiple fields and formats. Multi-source heterogeneous data refers to data sets from different sources with significant differences in format and structure. Specifically, in the context of substation site selection, this includes structured data such as power grid load data and geospatial information (e.g., topography, transportation networks, land use types), as well as unstructured or semi-structured data related to substation site selection, such as text reports (e.g., power planning documents, environmental impact assessment reports, policy documents), remote sensing images, infrared thermal imaging, and acoustic signatures. Remote sensing images can accurately present spatial information such as terrain slope, vegetation cover, and the distribution of surrounding buildings in the selected area, providing support for judging the terrain suitability of the substation site. Infrared thermal imaging can capture the regional ambient temperature distribution and the operating thermal status of existing power facilities in the surrounding area, assisting in identifying hidden safety hazards such as thermal interference. Acoustic signatures can monitor the noise level in the selected area, avoiding the impact of surrounding noise on substation operation and the lives of nearby residents. Text reports contain in-depth information such as regional power load planning, land use policies, and environmental protection standards, directly determining the compliance and feasibility of the site. These multi-dimensional data collectively constitute a complete data support system for substation site selection decisions.

[0004] Currently, AI-based site selection decision-making systems have become the core carrier of intelligent site selection technology. These systems process and analyze data using algorithms such as machine learning and deep learning to achieve intelligent evaluation and recommendation of site selection schemes. However, the development of existing AI-based site selection decision-making systems still faces significant technical bottlenecks. Most systems are limited by data processing capabilities and fusion technologies, and can only utilize single-modal data for site selection analysis. Some systems focus solely on grid load data, emphasizing the assessment of a site's power supply coverage and load-bearing capacity, while neglecting key factors such as geographic space and environmental constraints. Other systems rely solely on geographic information data, focusing on analyzing the site's terrain adaptability and transportation convenience, failing to fully integrate core indicators such as load demand, policy regulations, and environmental impact. Ultimately, they fail to achieve full fusion and in-depth analysis of multi-source heterogeneous data required for substation site selection, including text reports, remote sensing images, infrared thermal imaging, and acoustic signatures.

[0005] The aforementioned limitations directly affect the analytical accuracy and decision-making reliability of AI site selection decision-making systems, causing the system's output site selection schemes to often deviate from actual needs. This restricts the comprehensive and in-depth perception and analysis of the site selection environment, potentially leading to inaccurate site selection decisions. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a substation site selection method and system based on artificial intelligence, which aims to solve the problem that the existing substation site selection methods are not accurate enough.

[0007] One object of the present invention is to provide a substation site selection method based on artificial intelligence, the method comprising: Acquire multi-source heterogeneous raw data of the proposed site area, and perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions. A pre-trained cross-modal fusion neural network model is used to extract features and perform interactive analysis on a multimodal location dataset, generating a globally correlated feature vector of the location environment, specifically including: High-dimensional visual features of remote sensing images and infrared thermal images are extracted using convolutional neural networks or visual Transformers, spectral features of voiceprint signals are extracted using audio neural networks, and textual semantic features of text reports are extracted using pre-trained language models. A cross-modal attention mechanism module is constructed, using text semantic features as query vectors and visual and spectral features as key vectors to calculate the association weights between different modalities, thereby filtering out visual and auditory features that are strongly related to the location task and obtaining the fused global association feature vector. The global correlation feature vector is input into the multi-task decision head, and the comprehensive score of each candidate point is calculated through a weighted fusion strategy, and the final site selection decision result is output. Furthermore, in the aforementioned AI-based substation site selection method, the step of aligning and preprocessing the multi-source heterogeneous raw data with a spatiotemporal reference includes: The remote sensing image and the infrared thermal image are registered at the pixel level using geographic information system coordinates, and the voiceprint signal is mapped to the corresponding spatial coordinate point; Key geographic entities and constraints in text reports are extracted using natural language processing techniques and vectorized into text semantic features.

[0008] Furthermore, in the aforementioned AI-based substation site selection method, the step of extracting key geographic entities and constraints from the text report using natural language processing technology and vectorizing them into text semantic features includes: Construct a location selection knowledge graph based on a graph structure. The nodes of the knowledge graph include geographic entities, geological parameters and planning constraints, and the edges represent the spatial adjacency or logical constraint relationships between nodes. A pre-trained graph attention network is used to reason about the location knowledge graph, calculate the importance score of each node relative to the location target, and select key constraint nodes based on the importance score. The text fragments corresponding to the key constraint nodes are input into a bidirectional long short-term memory network to extract context semantic vectors. The context semantic vectors are then concatenated with the topology vectors output by the graph attention network to obtain the final text semantic features.

[0009] Furthermore, in the aforementioned AI-based substation site selection method, the step of constructing a cross-modal attention mechanism module, using text semantic features as query vectors and visual and spectral features as key vectors, calculates the correlation weights between different modalities to filter out visual and auditory features strongly related to the site selection task, and obtains a fused global correlation feature vector includes: Text semantic features are mapped to query vectors through a linear projection layer. The high-dimensional visual features of remote sensing images and infrared thermal images, as well as the spectral features of the voiceprint signal, are mapped into visual key vectors through independent linear projection layers. Visual value vector Auditory key vector and auditory value vector ; Calculate query vector With visual key vector The dot product similarity is used to obtain the text-visual attention map, and the query vector is calculated simultaneously. With auditory key vector The dot product similarity is used to obtain the text-auditory attention map; Normalize the text-visual attention map and the text-auditory attention map respectively to obtain the visual association weight matrix. Hearing association weight matrix ; With visual value vector As a query, with auditory key vector As the key, with auditory value vector As values, the guiding weights of visual features on auditory features are calculated to generate visually guided auditory enhancement features. ; Visual association weight matrix Acting on the visual value vector Weighted visual features are obtained The auditory association weight matrix Auditory enhancement features acting on visual guidance Weighted auditory features were obtained. ; Weighted visual features and weighted auditory features Channel-level concatenation is performed, and feature fusion is carried out through a multilayer perceptron to output a globally correlated feature vector.

[0010] Furthermore, in the aforementioned AI-based substation site selection method, the steps of inputting the global correlation feature vector into the multi-task decision head, calculating the comprehensive score of each candidate site through a weighted fusion strategy, and outputting the final site selection decision result include: The multi-task decision head includes a suitability classifier and a risk regressor, which output the suitability probability and potential risk value of the site selection, respectively. During the decision-making process, a conflict resolution algorithm based on evidence theory is introduced. When there is a logical conflict between the abnormal area shown by infrared thermal imaging and the safety description in the text report, the decision weight of the infrared mode is automatically increased, and a secondary verification is performed by combining the transient abnormal features in the voiceprint signal. Finally, a site selection report containing recommended coordinates, risk level and decision basis is generated.

[0011] Furthermore, in the aforementioned AI-based substation site selection method, the step of automatically increasing the decision weight of the infrared modality when a logical conflict arises between an abnormal area displayed by infrared thermal imaging and a safety description in a text report, involves introducing a conflict resolution algorithm based on evidence theory. The thermal anomaly features extracted from the infrared thermal image and the security constraint semantic features extracted from the text report are respectively regarded as two independent evidence bodies, with the infrared thermal image evidence body being the most important. Corresponding to the confidence level of thermal anomaly evidence, the text report evidence body Corresponding to the security description confidence evidence, both evidence bodies are mapped to a pre-defined location security identification framework. The identification framework It includes three identification elements: "safety", "risk", and "uncertainty". Calculate the two pieces of evidence , The basic probability assignment function, where The BPA function is calculated based on the temperature difference, area ratio, and duration of continuous occurrence of abnormal regions in the infrared thermal image. The BPA function is calculated based on the semantic similarity, frequency of occurrence of security constraint keywords in the text report, and importance scores of constraint nodes in the knowledge graph. The Dempster-Shafer synthesis rule is used to calculate the fusion probability of two pieces of evidence, and the probability is quantified by introducing a conflict coefficient. and The degree of logical conflict is determined. When the conflict coefficient is greater than or equal to the preset conflict threshold, it is determined that there is a logical conflict between the abnormal area displayed by the infrared thermal imaging and the security description in the text report. Using the conflict coefficient as input, and combining the hazard level weight of thermal anomaly risk in the substation site selection scenario, the weight adjustment increment of the infrared mode is calculated through a nonlinear mapping function, which is trained based on the review data of historical site selection conflict cases. Based on the adjusted infrared modal weights, the input feature weights of the suitability classifier and risk regressor in the multi-task decision head are recalculated. The adjusted weights are then applied to the high-dimensional visual features of infrared thermal imaging, while the weight ratio of the semantic features of the text report is simultaneously reduced, thus completing the feature weight calibration after conflict resolution.

[0012] Furthermore, in the aforementioned AI-based substation site selection method, the step of performing secondary verification by combining transient anomaly features in the acoustic signature signal to ultimately generate a site selection report containing recommended coordinates, risk level, and decision-making criteria includes: Transient anomaly features within a preset frequency band are extracted from the voiceprint signal. An algorithm combining wavelet packet decomposition and short-time Fourier transform is used to decompose the voiceprint signal at multiple scales. Transient signal segments with a duration of a preset time and an amplitude exceeding a preset multiple of the normal voiceprint baseline are selected as candidate sets of transient anomaly features. The candidate set of transient anomaly features is input into a pre-trained audio classification network. Combined with a sample library of typical interference sound patterns in the substation site selection scenario, the candidate set is classified and identified, and invalid interference transient features are eliminated to obtain effective transient anomaly features related to thermal anomalies. Calculate the spatial correlation and feature similarity between effective transient anomaly features and infrared thermal anomaly features. If the overlap between the spatial coordinates corresponding to the effective transient anomaly features and the infrared thermal anomaly region is greater than a threshold, and the feature similarity is greater than a threshold, then the effectiveness of the infrared thermal anomaly features is confirmed, and the current site selection risk assessment results are strengthened.

[0013] Another object of the present invention is to provide an artificial intelligence-based substation site selection system, the system comprising: The acquisition module is used to acquire multi-source heterogeneous raw data of the site selection area, and to perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions. The generation module is used to perform feature extraction and interaction analysis on multimodal location datasets using a pre-trained cross-modal fusion neural network model, generating a globally correlated feature vector of the location environment, specifically including: High-dimensional visual features of remote sensing images and infrared thermal images are extracted using convolutional neural networks or visual Transformers, spectral features of voiceprint signals are extracted using audio neural networks, and textual semantic features of text reports are extracted using pre-trained language models. A cross-modal attention mechanism module is constructed, using text semantic features as query vectors and visual and spectral features as key vectors to calculate the association weights between different modalities, thereby filtering out visual and auditory features that are strongly related to the location task and obtaining the fused global association feature vector. The location selection module is used to input the global correlation feature vector into the multi-task decision head, calculate the comprehensive score of each candidate point through a weighted fusion strategy, and output the final location selection decision result.

[0014] Another object of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0015] Another object of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0016] This invention acquires multi-source heterogeneous raw data of the proposed site selection area, performs spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data, and constructs a multimodal site selection dataset including remote sensing images, infrared thermal images, voiceprint signals, and text reports. A pre-trained cross-modal fusion neural network model is used to extract high-dimensional visual features, spectral features, and textual semantic features corresponding to each type of data. A cross-modal attention mechanism module is then constructed to calculate the correlation weights between different modalities, filter strongly correlated features, and generate a global correlation feature vector of the site selection environment. This global correlation feature vector is input into a multi-task decision head, and a weighted fusion strategy is used to calculate the comprehensive score of each candidate site, outputting the final site selection decision result. First, a unified preprocessing and standardization process is performed on the multi-source heterogeneous raw data involved in substation site selection to construct a multimodal dataset. Then, independent feature extraction is completed for each modality of data through multiple types of networks. Relying on a cross-modal attention mechanism, deep interactive fusion and effective filtering of multimodal features are achieved. Finally, the site selection result is output by combining a weighted score from a multi-task decision head. This breaks through the limitation of existing technologies that can only use single-modal data for site selection analysis, and achieves full fusion and in-depth mining of multi-source heterogeneous data such as remote sensing images, infrared thermal imaging, acoustic signals, and text reports. It solves the problems of existing AI site selection decision systems being unable to fully integrate multi-source heterogeneous data, resulting in incomplete site selection environment perception and analysis, low analysis accuracy, poor decision reliability, and deviations between site selection results and actual needs. Attached Figure Description

[0017] Figure 1 This is a flowchart of a substation site selection method based on artificial intelligence in the first embodiment of the present invention; Figure 2 This is a structural block diagram of an artificial intelligence-based substation site selection system according to the third embodiment of the present invention.

[0018] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0019] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0020] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1 Please see Figure 1 The figure shows a substation site selection method based on artificial intelligence in the first embodiment of the present invention, the method including steps S10 to S12.

[0023] Step S10: Obtain multi-source heterogeneous raw data of the area to be selected, and perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions.

[0024] First, it is necessary to collect various types of raw data from different sources and with inconsistent formats (i.e., multi-source heterogeneous raw data) from the site selection area. This raw data forms the basis for subsequent site selection work. The multi-source heterogeneous raw data should cover at least four key types of data: first, remote sensing images reflecting the geospatial characteristics of the area (such as topography, landforms, and land use); second, infrared thermal images reflecting the operating status of equipment or the distribution of surface thermal anomalies (such as equipment overheating and surface hotspots); third, voiceprint signals reflecting the environmental sound field characteristics of the area (such as noise distribution and changes in sound intensity); and fourth, text reports containing the results of geological surveys in the area (such as geological structure and soil conditions) or planning restrictions (such as land use rights and construction standards).

[0025] Next, two key processes need to be performed on the collected multi-source heterogeneous raw data: spatiotemporal benchmark alignment and preprocessing. Spatiotemporal benchmark alignment is to ensure that all data are consistent in time and space dimensions, and to avoid the inability of data to be effectively integrated and used due to differences in spatiotemporal benchmarks. Preprocessing is to perform operations such as cleaning, denoising, and standardizing the data to remove invalid information, optimize data quality, and prepare for subsequent dataset construction. Finally, through the above processing, all valid data are integrated to construct a multimodal location dataset containing various data formats. For example, the steps of aligning and preprocessing multi-source heterogeneous raw data with a spatiotemporal reference include: The remote sensing image and the infrared thermal image are registered at the pixel level using geographic information system coordinates, and the voiceprint signal is mapped to the corresponding spatial coordinate point; Key geographic entities and constraints in text reports are extracted using natural language processing techniques and vectorized into text semantic features.

[0026] Among them, spatiotemporal benchmark alignment refers to unifying different data under the same time and space reference standard to ensure that the data can be correlated and fused; preprocessing refers to cleaning, optimizing and extracting features from the raw data to meet the needs of subsequent dataset construction and analysis; pixel-level registration refers to accurately matching the pixel positions of two or more images to eliminate spatial offset; and text semantic feature vectorization refers to converting key information in the text into a vector form that can be recognized and calculated by computers.

[0027] First, for three types of non-textual data—remote sensing images, infrared thermal images, and voiceprint signals—a common Geographic Information System (GIS) coordinate system (such as the WGS84 coordinate system, a commonly used benchmark in typical site selection scenarios, which can be adjusted to the local coordinate system according to the actual site selection area) is used as a unified spatial reference benchmark. First, pixel-level registration is performed between the remote sensing images and the infrared thermal images. For example, the SIFT algorithm (Scale Invariant Feature Transform algorithm, a conventional image registration algorithm) is used to extract key feature points from the two images (such as terrain inflection points in the remote sensing image and thermal anomaly boundary points in the infrared thermal image). Then, the RANSAC algorithm (Random Sample Consensus algorithm) is used to remove mismatched feature points. Finally, an affine transformation is used to map the pixel coordinates of the infrared thermal image to the GIS coordinate system of the remote sensing image, achieving precise pixel-level alignment between the two and ensuring that pixel information at the same spatial location can correspond (such as precise matching of remote sensing terrain information and infrared thermal anomaly information for a certain plot). Simultaneously, spatial coordinate mapping processing is performed on the voiceprint signal. Specifically, this involves first acquiring the original spatial location information (such as GPS positioning data from the acquisition device) during voiceprint signal acquisition, and then converting it into data that can be mapped to remote sensing images and infrared thermal imaging. Figure 1 The GIS coordinates are then used to decompose the continuous voiceprint signal (such as the noise intensity signal of a certain area) into discrete signal points through linear interpolation (a conventional signal spatial mapping method), and accurately map them to the corresponding spatial coordinate points. This makes the voiceprint signal and image data completely unified in the spatial dimension, realizing the spatial synergy of "image + signal".

[0028] Secondly, for text report data, preprocessing and semantic feature extraction can be performed. Conventional Natural Language Processing (NLP) techniques can be used to process text reports containing geological surveys and planning restrictions throughout the entire process. For example, firstly, text cleaning techniques are used to remove invalid information (such as redundant spaces, irrelevant comments, and duplicate content) from the text. Then, word segmentation algorithms (such as jieba segmentation, a conventional Chinese word segmentation method) are used to break the text down into independent words. Subsequently, named entity recognition (NER) algorithms (such as the BERT model) are used to accurately extract key geographic entities (such as "XX") from the text. The key geographic entities and constraints are extracted, including "XX region" and "sandy loam" in the phrase "regional geology is sandy loam," and "building height" and "20 meters" in the phrase "building height not exceeding 20 meters in planning restrictions." These constraints include geological constraints such as "construction is prohibited below the groundwater level in this area" and planning constraints such as "industrial construction is prohibited in this area, which is an ecological protection zone." After extraction, word embedding technology (such as the Word2Vec model and conventional text vectorization methods) is used to transform the extracted key geographic entities and constraints into fixed-dimensional text semantic feature vectors (e.g., transforming "sandy loam" into a vector form of [0.123, 0.456, ...]). This transforms the text data from "natural language" into "computer-computable feature vectors," thereby achieving format unification between the text data and the aforementioned image and signal data.

[0029] Furthermore, in some optional embodiments of the present invention, the step of extracting key geographic entities and constraints from the text report using natural language processing technology and vectorizing them into text semantic features includes: Construct a location selection knowledge graph based on a graph structure. The nodes of the knowledge graph include geographic entities, geological parameters and planning constraints, and the edges represent the spatial adjacency or logical constraint relationships between nodes. A pre-trained graph attention network is used to reason about the location knowledge graph, calculate the importance score of each node relative to the location target, and select key constraint nodes based on the importance score. The text fragments corresponding to the key constraint nodes are input into a bidirectional long short-term memory network to extract context semantic vectors. The context semantic vectors are then concatenated with the topology vectors output by the graph attention network to obtain the final text semantic features.

[0030] Among them, Natural Language Processing (NLP) technology refers to a series of technologies for processing, analyzing, and understanding natural language text, used here for key information extraction and semantic vectorization in text reports; Geographic entities refer to geographically related objects in the text (such as a region name, geological type, topography, etc.); Constraints refer to the limiting requirements in the text that affect site selection (such as geological constraints, planning constraints, environmental constraints, etc.); Text semantic features refer to the transformation of text information into a computer-recognizable and computable vector form for subsequent data fusion and analysis; Knowledge graph is a semantic network based on a graph structure, composed of nodes and edges, where nodes represent entities or concepts and edges represent the relationships between nodes; Graph structure is a topological structure composed of nodes (V) and edges (E), used to describe the relationships between entities; Graph Attention Network (GAT) is a basic... Graph neural networks based on attention mechanisms can calculate the importance of relationships between nodes through attention weights, enabling reasoning and feature extraction of graph-structured data; pre-trained models: models that have been pre-trained on large-scale datasets and can be directly transferred to adapt to specific tasks. Here, pre-trained graph attention networks can improve reasoning efficiency and accuracy; Bidirectional Long Short-Term Memory (Bi-LSTM): an improved recurrent neural network (RNN) that can simultaneously capture the contextual information of text (forward and backward) and is good at extracting semantic features of sequential text; Contextual semantic vector: output by Bi-LSTM, reflecting the contextual semantic information of text segments; Topological structure vector: output by GAT, reflecting the relationship features between nodes in the knowledge graph; Feature concatenation: fusing two vectors of different dimensions into a comprehensive vector, retaining two types of feature information.

[0031] First, a graph-based knowledge graph for site selection can be constructed using a "text parsing - node extraction - edge construction" architecture. The text report is initially parsed (using word segmentation and part-of-speech tagging techniques, such as jieba segmentation and StanfordPOS tagger), extracting three types of core nodes (V): geographic entity nodes (e.g., "XX Industrial Park", "sandy loam", "groundwater level"), geological parameter nodes (e.g., "soil bearing capacity ≥150kPa", "groundwater level depth ≤5m"), and planning restriction nodes (e.g., "industrial construction prohibited", "building height ≤20m"). Subsequently, based on the semantic associations in the text, edges (E) between nodes are constructed. The edge types are divided into two categories: one is spatial adjacency (such as the "spatial overlap" relationship between "XX Industrial Park" and "groundwater level"), and the other is logical constraint relationship (such as the "satisfaction" relationship between "sandy loam" and "soil bearing capacity ≥150kPa", and the "constraint" relationship between "prohibited industrial construction" and "XX Industrial Park"). Finally, a complete site selection knowledge graph is formed. Secondly, a pre-trained Graph Attention Network (GAT) is used to reason about the location selection knowledge graph and filter key constraint nodes. Specifically: first, a pre-trained GAT model (such as a GAT model pre-trained based on PubMedBERT) is used as input to the constructed location selection knowledge graph, and the model calculates the value of each node through the attention mechanism. Importance score relative to site selection objectives , For example, as one implementation method for filtering key constraint nodes: Each node The original features (such as word vectors after text segmentation) represent ( (for feature dimensions) For nodes Each neighbor node Calculate the attention coefficients of both. ,reflect right The importance is expressed by the formula:

[0032] in, The weight matrix ( (the transformed feature dimensions). Let be the attention weight vector, and LeakyReLU be the activation function. For nodes The set of neighboring nodes, " indicates vector concatenation; Based on attention coefficient The features of neighboring nodes are weighted and summed to obtain the node. Updated features The formula is: , where σ is the sigmoid activation function; After multiple iterations of the GAT model (typically 3-5 iterations), the final feature vector of each node is output. It is mapped to a single importance score through a fully connected layer. The formula is: ; in For the weights of the fully connected layer, For bias terms; Key constraint node selection: Set a threshold τ, and select nodes based on their importance scores. When the value is greater than or equal to τ, it is determined to be a critical constraint node, and the set of critical constraint nodes is obtained after filtering. V' ; Finally, the text segments corresponding to the key constraint nodes are input into a Bidirectional Long Short-Term Memory (Bi-LSTM) network, and the features are fused to obtain the final text semantic features. A "Bi-LSTM feature extraction + GAT topology feature fusion" architecture can be adopted. First, the text segments corresponding to each key constraint node are extracted, and these text segments are converted into word vector sequences (using the Word2Vec model to convert each word into a word vector sequence). 3D word vectors, resulting in a sequence of word vectors. (where t is the length of the text segment). The word vector sequence is then input into the Bi-LSTM model, from the forward direction (from...) arrive ) and reverse (from arrive Capture contextual semantic information to obtain the positive hidden state. and reverse hidden state The positive hidden state and reverse hidden state By concatenating the vectors, we obtain the contextual semantic vectors of each text segment. Simultaneously, the topological structure vectors of the key constraint nodes output by the GAT model in the second step (i.e., the node features after GAT update) are extracted. ); Finally, the context semantic vector With topological structure vector By concatenating the features, we obtain the final text semantic feature vector corresponding to each key constraint node. The set of text semantic feature vectors of all key constraint nodes is the final text semantic feature of the entire text report.

[0033] Step S11: Use a pre-trained cross-modal fusion neural network model to perform feature extraction and interactive analysis on the multimodal location dataset to generate a global correlation feature vector of the location environment.

[0034] Among them, a cross-modal fusion neural network model that has completed the pre-training process is used to carry out two core operations on the site selection dataset containing multiple modal information. The first is feature extraction, which is to accurately extract representative core information that can reflect the key conditions of site selection from the site selection data of various modalities. The second is interactive analysis, which is to conduct deep correlation and cross-analysis on the features extracted from different modal data, and explore the intrinsic connections and mutual influences between the features of various modalities. Finally, through this series of operations, a global correlation feature vector that can comprehensively and systematically reflect the overall situation of the site selection environment is generated.

[0035] First, high-dimensional features of a single modality are extracted using a "modality-specific adaptive model" architecture. High-dimensional features are extracted for each of the four modalities. The specific implementation details are as follows: High-dimensional visual feature extraction from remote sensing images and infrared thermal images: targeting (Remote sensing images) and (Infrared thermal imaging) If the image is rich in detail and has key local features (such as terrain inflection points in remote sensing images or thermal anomaly boundaries in infrared thermal imaging), a convolutional neural network (CNN) is used, with ResNet50 as an example model. Through 50 layers of convolution and pooling operations, local features and spatial textures of the image are extracted, and a high-dimensional visual feature vector is finally output. If the global correlation features of the image are more important (such as the overall regional layout of remote sensing images or the global distribution of thermal anomalies in infrared thermal imaging), a visual transformer (ViT) is used, with ViT-B / 16 as an example model. The image is divided into 16×16 image blocks, and the global correlation between image blocks is captured through a self-attention mechanism. The high-dimensional visual feature vector with the same dimension as that extracted by CNN is then concatenated to obtain a merged visual feature vector, integrating the core features of the two visual modalities.

[0036] Spectral feature extraction of voiceprint signals: targeting For voiceprint signals, the Mel-CNN model can be used. First, the time-domain voiceprint signal (continuous sound waveform) is converted into a frequency-domain Mel spectrogram using Mel frequency cepstral coefficients (MFCC) (mapping the sound frequency to the Mel scale that is sensitive to human hearing, which is more in line with the sound field feature analysis). Then, the Mel spectrogram is processed by convolution and pooling through CNN to extract spectral features (such as noise intensity peaks, frequency distribution ranges, sound field uniformity, etc.), and finally outputs the spectral feature vector of the voiceprint signal. If the voiceprint signal has temporal variations (such as differences in noise intensity at different times), an LSTM layer can be added after the CNN to capture the temporal correlation of spectral features and optimize the representation ability.

[0037] Text semantic feature extraction from text reports: targeting (Text report) can be processed using a conventional pre-trained language model, with BERT-base as an example. The text report is segmented and encoded (converted into a token sequence that the model can recognize), and then input into the encoder of the BERT model. The self-attention mechanism captures the contextual semantic relationships of the text and extracts core information such as geographical entities, geological parameters, and planning constraints from the text, finally outputting a text semantic feature vector. To adapt to the query requirements of subsequent cross-modal attention mechanisms, a fully connected layer is used to map the text semantic feature vector to an intermediate dimension that is unified with visual features and spectral features, resulting in a text semantic feature vector adapted to the query vector.

[0038] Next, cross-modal attention interaction filtering and global correlation feature vector generation are performed. Relying on the constructed cross-modal attention mechanism module, guided by text semantic features, visual and audio features strongly related to the location selection task are filtered, and finally fused to generate a global correlation feature vector. The specific implementation method is as follows: Cross-modal attention mechanism module construction: The module as a whole adopts a "query-key-value" approach. The classic attention architecture of "" clearly defines the role of each modality feature—using the processed text semantic feature vector. as query vector Its core function is to guide the selection of features that are "strongly correlated with textual constraints and geographic entities"; using the merged visual feature vector. and spectral eigenvectors as a key vector Sum value vector ,in , The key vector and value vector use the same feature vector to ensure that the filtered features are consistent with the original features, and only redundant information is filtered out. The module contains three core steps: attention weight calculation, normalization, and weighted summation, which enable the filtering of strongly correlated features.

[0039] Association weight calculation and feature selection: The association weights between Q and K are calculated through an attention mechanism to characterize the degree of association between each dimension of visual features and spectral features and textual semantic features (location requirements), resulting in an association weight matrix. The weight values ​​range from [0,1]. The higher the weight, the stronger the correlation between the corresponding feature dimension and the location task. The weight matrix W and the value vector V are then summed using weighted averages to obtain the filtered fused feature vector. This vector retains only the visual and spectral features that are strongly correlated with the semantic features of the text (site selection requirements), and removes redundant and irrelevant features (such as background details in remote sensing images that are irrelevant to site selection, and meaningless noise interference in the voiceprint signal).

[0040] Global correlation feature vector generation: This involves generating the filtered and fused feature vectors. Compared with the semantic feature vector of the original text The features are then concatenated and fused to ensure that the final features include both the selected strongly relevant visual and audio features and retain the core textual constraints. The formula is as follows: ,in This is the final generated global correlation feature vector of the location environment; to optimize feature representation capabilities, a fully connected layer can be used to... Mapping to a unified target dimension (usually set to 1024), while using a dropout layer to avoid overfitting, the final output is a standardized global correlation feature vector.

[0041] Step S12: Input the global correlation feature vector into the multi-task decision head, calculate the comprehensive score of each candidate point through a weighted fusion strategy, and output the final site selection decision result.

[0042] After generating a globally unified feature vector, based on multiple pre-set candidate substation locations in the area to be selected, the local and global feature vectors corresponding to each location are extracted separately. The feature vector of each candidate location is then input into the multi-task decision head at the end of the model. The multi-task decision head can simultaneously complete the dual tasks of site suitability assessment and risk level assessment. Combining a pre-trained weighted fusion strategy, it assigns weighted scores to multiple assessment dimensions for each location, including geographical adaptability, geological safety, environmental stability, planning compliance, and risk level. The comprehensive site selection score for each candidate location is obtained by combining the scores of all dimensions. Based on the comprehensive scores, the locations are sorted from highest to lowest. Combined with pre-set site selection rules, the optimal substation site selection decision result is finally output, completing the entire intelligent site selection process.

[0043] Furthermore, the multi-task decision head includes a suitability classifier and a risk regressor, which respectively output the suitability probability and potential risk value of the site selection; During the decision-making process, a conflict resolution algorithm based on evidence theory is introduced. When there is a logical conflict between the abnormal area shown by infrared thermal imaging and the safety description in the text report, the decision weight of the infrared mode is automatically increased, and a secondary verification is performed by combining the transient abnormal features in the voiceprint signal. Finally, a site selection report containing recommended coordinates, risk level and decision basis is generated.

[0044] The multi-task decision head internally comprises two independent functional modules: a suitability classifier and a risk regression analyzer. These two modules perform their respective functions in parallel to complete site selection-related tasks. The suitability classifier employs a softmax classification structure, using substation construction adaptation standards as training criteria. It classifies and judges the overall suitability of each candidate site, outputting a site suitability probability ranging from 0 to 1. A higher probability indicates a better fit for substation construction requirements. The risk regression analyzer uses a fully connected regression network structure, trained with a large number of historical site selection risk samples. It quantitatively assesses the geological risks, environmental risks, equipment safety risks, and planning compliance risks of candidate sites, outputting specific potential risk values. Higher risk values ​​indicate greater safety hazards at the site.

[0045] Furthermore, in actual site selection scenarios, contradictory multimodal data information is highly likely to arise. A typical example is when infrared thermal imaging captures risk areas with abnormal surface temperatures or hidden heat hazards, but the written planning report and geological report describe the safety status of these areas as compliant and safe. This creates a logical conflict between the two types of information, leading to distorted decisions if made directly. To address this issue, this invention introduces a conflict resolution algorithm based on evidence theory. When a logical conflict is detected between an abnormal area detected by infrared thermal imaging and the safety description in the text, the system automatically increases the decision weight of the infrared modality. Since infrared thermal imaging data is real-time, real-scene detection data, it is more accurate and timely. Text reports, on the other hand, suffer from data update delays and vague descriptions. Increasing the weight of the infrared modality prioritizes decisions based on real-scene risk characteristics.

[0046] After resolving modal weight conflicts, a second cross-validation of infrared thermal anomaly risks is performed based on the transient anomaly characteristics of the acoustic signature signal in the candidate site area. This further verifies the authenticity and effectiveness of the risks and eliminates misjudgment interference from single-modal detection. Following this validation, the suitability probability, quantified risk value, modal feature matching results, and conflict resolution results of each candidate site are integrated to generate standardized site selection decision results. The final output is a professional site selection report containing accurately recommended latitude and longitude coordinates, clear risk level classifications, and complete supporting evidence for the decision, thus completing the overall site selection decision output.

[0047] For example, the step of automatically increasing the decision weight of the infrared modality when there is a logical conflict between the abnormal area displayed by the infrared thermal imaging and the security description in the text report, by introducing a conflict resolution algorithm based on evidence theory, includes: The thermal anomaly features extracted from the infrared thermal image and the security constraint semantic features extracted from the text report are respectively regarded as two independent evidence bodies, with the infrared thermal image evidence body being the most important. Corresponding to the confidence level of thermal anomaly evidence, the text report evidence body Corresponding to the security description confidence evidence, both evidence bodies are mapped to a pre-defined location security identification framework. The identification framework Θ includes three identification elements: "security", "risk", and "uncertainty". Calculate the two pieces of evidence , The basic probability assignment function, where The BPA function is calculated based on the temperature difference, area ratio, and duration of continuous occurrence of abnormal regions in the infrared thermal image. The BPA function is calculated based on the semantic similarity, frequency of occurrence of security constraint keywords in the text report, and importance scores of constraint nodes in the knowledge graph. The Dempster-Shafer synthesis rule is used to calculate the fusion probability of two pieces of evidence, and the probability is quantified by introducing a conflict coefficient. and The degree of logical conflict is determined. When the conflict coefficient is greater than or equal to the preset conflict threshold, it is determined that there is a logical conflict between the abnormal area displayed by the infrared thermal imaging and the security description in the text report. Using the conflict coefficient as input, and combining the hazard level weight of thermal anomaly risk in the substation site selection scenario, the weight adjustment increment of the infrared mode is calculated through a nonlinear mapping function, which is trained based on the review data of historical site selection conflict cases. Based on the adjusted infrared modal weights, the input feature weights of the suitability classifier and risk regressor in the multi-task decision head are recalculated. The adjusted weights are then applied to the high-dimensional visual features of infrared thermal imaging, while the weight ratio of the semantic features of the text report is reduced simultaneously, thus completing the feature weight calibration after conflict resolution.

[0048] To address the information conflict between infrared thermal imaging and text reporting modalities, two independent evidence bodies are constructed to characterize decision-making evidence information from these two data sources. The infrared thermal imaging evidence body... Specifically designed for regional thermal anomaly confidence level evidence, used to characterize the credibility of potential risks such as thermal hazards, equipment malfunctions, and geological disturbances causing heat generation in the proposed site area; text report evidence. Specifically corresponding to the confidence level evidence of regional security descriptions, this is used to characterize the credibility of textual data regarding regional security compliance and the absence of construction risks. Both types of evidence are uniformly mapped to a pre-defined site selection security identification framework. This framework is specifically designed for substation site selection scenarios. It contains three mutually exclusive and complete identification elements: safety, risk, and uncertainty. All evidence assessment results are classified into one of these three elements, ensuring a consistent standard for evidence evaluation.

[0049] Next, the basic probability allocation functions for the two pieces of evidence are calculated separately. These functions quantify the probability of each piece of evidence supporting different outcomes within the identification framework. Different calculation dimensions are used for the two pieces of evidence to ensure the accuracy of evidence quantification. For the infrared thermal imaging evidence... Its BPA function mainly relies on three pieces of real-world data for calculation: the temperature difference between abnormal and normal areas in the infrared image, the area ratio of the abnormal area to the overall detection area, and the duration of continuous occurrence of thermal anomalies. These three indicators are weighted and fitted to calculate the final output. The basic probability distributions for three types of outcomes: safety, risk, and uncertainty. For textual report evidence... Its BPA function is calculated based on three textual feature data: the semantic similarity between the textual security constraint keywords and the standard security specifications, the frequency of occurrence of core security clauses in the report, and the importance score of the corresponding security constraint node in the knowledge graph. These are obtained through weighted fusion calculation. The basic probability distribution of the three types of recognition results.

[0050] Subsequently, the classic DS evidence synthesis rules were applied to... , The basic probability allocation results of the two pieces of evidence are fused together. During the fusion process, the evidence conflict coefficient is calculated simultaneously. This coefficient is the core indicator for quantifying the degree of logical contradiction between the two types of evidence. The larger the value, the more significant the contradiction between the infrared thermal imaging measured risk evidence and the textual security description evidence. A fixed conflict threshold is pre-set based on a large amount of historical site selection data. When the calculated conflict coefficient is greater than or equal to the preset conflict threshold, the system accurately determines that there is a logical conflict between the infrared anomaly and the textual security description, triggering a weight adjustment mechanism. If the conflict coefficient is less than the threshold, it is determined that there is no obvious conflict, and the original modality weights are used directly for decision-making.

[0051] Using the quantified conflict coefficient as the core input variable, and combining it with the hazard level weights corresponding to thermal anomaly risks in substation site selection scenarios, a dedicated nonlinear mapping function is introduced to complete the weight increment calculation. This nonlinear mapping function is not a fixed formula, but a dedicated function trained using machine learning by collecting massive amounts of historical site selection conflict cases, thermal anomaly risk accident cases, and modal decision deviation review data. It can adapt to the weight adjustment needs of different conflict degrees and risk levels, and can dynamically output precise infrared modal weight adjustment increments based on the K value. The higher the conflict degree, the greater the increase in infrared modal weight.

[0052] Finally, based on the calculated weight adjustment increments, the input feature weights of the multi-task decision head are globally corrected. The weight ratio of high-dimensional visual features from infrared thermal imaging is increased in the decision-making process, while the weight ratio of semantic features from text reports is simultaneously reduced to mitigate the interference of lagging text information. The calibrated weight parameters are then applied to the modal branches of the globally associated feature vectors, updating the input weight ratios of the suitability classifier and the risk regressor. This completely eliminates logical conflicts in the multimodal data, achieving adaptive conflict resolution and ensuring that the decision results closely reflect the real-world environment.

[0053] For example, the step of performing secondary verification by combining transient anomaly features in the voiceprint signal, and finally generating a site selection report containing recommended coordinates, risk level, and decision-making basis includes: Transient anomaly features within a preset frequency band are extracted from the voiceprint signal. An algorithm combining wavelet packet decomposition and short-time Fourier transform is used to decompose the voiceprint signal at multiple scales. Transient signal segments with a duration of a preset time and an amplitude exceeding a preset multiple of the normal voiceprint baseline are selected as candidate sets of transient anomaly features. The candidate set of transient anomaly features is input into a pre-trained audio classification network. Combined with a sample library of typical interference sound patterns in the substation site selection scenario, the candidate set is classified and identified, and invalid interference transient features are eliminated to obtain effective transient anomaly features related to thermal anomalies. Calculate the spatial correlation and feature similarity between effective transient anomaly features and infrared thermal anomaly features. If the overlap between the spatial coordinates corresponding to the effective transient anomaly features and the infrared thermal anomaly region is greater than a threshold, and the feature similarity is greater than a threshold, then the effectiveness of the infrared thermal anomaly features is confirmed, and the current site selection risk assessment results are strengthened.

[0054] First, for the collected raw environmental voiceprint signals, a fusion algorithm combining wavelet packet decomposition and short-time Fourier transform is used to conduct multi-scale feature analysis, taking into account the ability to capture detailed features in both the time and frequency domains. First, the voiceprint signal is segmented into frames using short-time Fourier transform to extract the time-domain amplitude changes and basic spectral features. Then, multi-level wavelet packet decomposition is used to further subdivide and analyze the voiceprint signals in different frequency bands, uncovering the differentiated features of high-frequency transients and low-frequency steady-state signals. Pre-defined voiceprint anomaly judgment criteria are established, filtering out transient signal segments with a duration within a preset time interval and a signal amplitude exceeding a preset multiple of the normal voiceprint baseline. All signal segments meeting the basic anomaly conditions are aggregated to form an initial transient anomaly feature candidate set, initially eliminating invalid signals such as steady-state environmental noise.

[0055] Subsequently, a pre-trained audio classification network adapted to substation site selection scenarios was built. Simultaneously, a sample library containing interference soundprints from typical scenarios such as wind noise, rain noise, vehicle traffic, bird calls, and normal equipment operating noise was constructed for accurate identification of abnormal features. All transient anomaly feature candidate sets obtained from the initial screening were input into the audio classification network. The network classifies and identifies each transient signal in the candidate set through feature matching and similarity comparison, accurately distinguishing between environmental interference noise and genuine abnormal sound field signals, automatically eliminating various invalid interference transient features, and ultimately selecting effective transient anomaly features highly correlated with thermal anomaly risks such as geological disturbances, underground hazards, and abnormal equipment heating and noise.

[0056] Finally, the spatial geographic coordinates corresponding to the effective transient anomaly features are accurately extracted. Simultaneously, the spatial coordinate range and feature parameters of the thermal anomaly area in the infrared thermal imaging are located. The spatial overlap and feature similarity of the two types of anomaly features are calculated respectively. Pre-set thresholds for spatial overlap and feature similarity are established. When the spatial overlap between the effective transient anomaly feature and the infrared thermal anomaly area exceeds the preset threshold, and the matching similarity of the two types of features also exceeds the preset threshold, the acoustic anomaly and thermal anomaly are determined to be caused by the same source of real risk, thoroughly confirming the validity of the infrared thermal anomaly features. Based on the real risk hazards verified by dual modality, the risk assessment level of the current candidate sites is strengthened, and the suitability score results are revised. Finally, the coordinate information, suitability score, risk level, modal detection basis, conflict resolution process, and secondary verification results of all candidate sites are integrated to generate a complete, well-founded, and accurate final substation site selection report, outputting the optimal site selection decision result.

[0057] In summary, the substation site selection method based on artificial intelligence in the above embodiments of the present invention acquires multi-source heterogeneous raw data of the area to be selected, performs spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data, and constructs a multimodal site selection dataset including remote sensing images, infrared thermal images, acoustic signals, and text reports; adopts a pre-trained cross-modal fusion neural network model to extract high-dimensional visual features, spectral features, and text semantic features corresponding to various types of data, and then calculates the correlation weights between different modalities by constructing a cross-modal attention mechanism module, filters strongly correlated features, and generates a global correlation feature vector of the site selection environment; inputs the global correlation feature vector into a multi-task decision head, calculates the comprehensive score of each candidate site through a weighted fusion strategy, and outputs the final site selection decision result. First, the multi-source heterogeneous raw data involved in substation site selection are uniformly preprocessed and standardized to construct a multimodal dataset. Then, independent feature extraction of each modality is completed through multiple types of networks. Relying on a cross-modal attention mechanism, deep interactive fusion and effective filtering of multimodal features are achieved. Finally, the site selection result is output by combining a weighted score from a multi-task decision head. This breaks through the limitation of existing technologies that can only use single-modal data for site selection analysis, and realizes the full fusion and in-depth mining of multi-source heterogeneous data such as remote sensing images, infrared thermal imaging, acoustic signals, and text reports. It solves the problems of existing AI site selection decision systems being unable to fully integrate multi-source heterogeneous data, resulting in incomplete site selection environment perception analysis, low analysis accuracy, poor decision reliability, and deviations between site selection results and actual needs.

[0058] Example 2 This embodiment also proposes an AI-based substation site selection method. The difference between the AI-based substation site selection method in this embodiment and the AI-based substation site selection method in Embodiment 1 is as follows: The steps of constructing the cross-modal attention mechanism module, which uses text semantic features as query vectors and visual and spectral features as key vectors, to calculate the association weights between different modalities, thereby filtering out visual and auditory features strongly related to the location selection task, and obtaining the fused global association feature vector, include: Text semantic features are mapped to query vectors through a linear projection layer. The high-dimensional visual features of remote sensing images and infrared thermal images, as well as the spectral features of the voiceprint signal, are mapped into visual key vectors through independent linear projection layers. Visual value vector Auditory key vector and auditory value vector ; Calculate query vector With visual key vector The dot product similarity is used to obtain the text-visual attention map, and the query vector is calculated simultaneously. With auditory key vector The dot product similarity is used to obtain the text-auditory attention map; Normalize the text-visual attention map and the text-auditory attention map respectively to obtain the visual association weight matrix. Hearing association weight matrix ; With visual value vector As a query, with auditory key vector As the key, with auditory value vector As values, the guiding weights of visual features on auditory features are calculated to generate visually guided auditory enhancement features. ; Visual association weight matrix Acting on the visual value vector Weighted visual features are obtained The auditory association weight matrix Auditory enhancement features acting on visual guidance Weighted auditory features were obtained. ; Weighted visual features and weighted auditory features Channel-level concatenation is performed, and feature fusion is carried out through a multilayer perceptron to output a globally correlated feature vector.

[0059] First, linear projection mapping of multimodal features is performed to generate query, key, and value vectors required for attention calculation. Linear dimension transformations are then applied to the preprocessed text semantic features, visual high-dimensional features, and voiceprint spectrum features to unify feature dimensions and generate vectors specifically for attention calculation. For example, fixed-dimensional text semantic features are input into an independent linear projection layer, and standardized query vectors are obtained through weight matrix operations. To address the high-dimensional visual features resulting from the fusion of remote sensing images and infrared thermal images, a dedicated independent linear projection layer is used to map and generate visual key vectors. Visual value vector Based on the spectral features extracted from the voiceprint signal, another independent linear projection layer is applied to generate auditory key vectors. and auditory value vector Each linear projection layer is a fixed-weight layer that has been trained on the model, which ensures that the dimensions of different modal vectors are consistent and meets the dimension matching requirements for attention dot product calculation.

[0060] Next, cross-modal attention maps are calculated modally to achieve semantic guidance of text based on visual and auditory features. Vector dot product similarity is used to calculate the correlation between the query vector and various key vectors. The query vector corresponding to the text is then... With visual key vector A dot-wise product operation is performed to quantify the correlation between textual planning constraints, geological constraints, and geographic visual features, generating a text-visual attention map. The weight value of each pixel in the attention map corresponds to the correlation strength between different visual regions and the site selection task. Simultaneously, the query vector is... With auditory key vector Dot product similarity calculation is performed to quantify the correlation between text security constraints, environmental constraints and sound field anomaly features, and a text-auditory attention map is generated to characterize the contribution of different voiceprint features to site selection risk assessment.

[0061] Then, to address the issues of uneven distribution and excessive differences in initial attention weight values, a softmax normalization function was used to globally normalize the text-visual attention map and the text-auditory attention map, respectively. After normalization, the attention weight values ​​of all regions are mapped to the interval between 0 and 1, and the sum of all weights under the same modality is 1. Finally, standardized visual association weight matrices and auditory association weight matrices are obtained, accurately quantifying the location association priority of each visual region and each audio feature.

[0062] Next, auditory feature enhancement is performed, using visual value vectors. As a query, auditory key vector As a key, the auditory value vector As a value, the guiding weight of visual features on auditory features is calculated through an attention mechanism (for example, if visual features indicate that a certain area is woodland, the model will focus on the natural voiceprint features of that area and weaken artificial noise features), thereby generating visually guided auditory enhancement features. (Optimized auditory features that incorporate visual guidance information make the auditory features more aligned with site selection requirements).

[0063] Subsequently, the features are weighted, and the visual association weight matrix is ​​calculated. Acting on the visual value vector (That is, by multiplying each visual feature with its corresponding weight using matrix multiplication), the weighted visual features are obtained. (Highlight visual features strongly relevant to the site selection task and downplay irrelevant visual features); simultaneously, incorporate auditory association weighting matrices. Auditory enhancement features acting on visual guidance Weighted auditory features were obtained. (Highlight auditory features that are strongly related to the site selection task).

[0064] Finally, feature fusion is completed, combining the weighted visual features. and weighted auditory features Perform channel-level concatenation (that is, merge the dimensions of two feature vectors, example: It is a 100-dimensional vector. The 50-dimensional vector is concatenated to obtain a 150-dimensional vector. Then, it is fused using a multilayer perceptron (a model composed of multiple neural networks that can perform nonlinear transformations and deep fusion on the concatenated features; for example, a simple neural network containing an input layer, hidden layer, and output layer, where the hidden layer uses the ReLU activation function to extract and optimize the concatenated features). The final output is a global associated feature vector (a comprehensive feature vector that integrates text semantics, key visual features, and key auditory features, which can be directly used for model training or decision-making in location selection tasks; for example, the global associated feature vector can be represented as [0.15, 0.4, 0.25,...], with each dimension corresponding to the weight of different fused features).

[0065] In summary, the substation site selection method based on artificial intelligence in the above embodiments of the present invention acquires multi-source heterogeneous raw data of the area to be selected, performs spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data, and constructs a multimodal site selection dataset including remote sensing images, infrared thermal images, acoustic signals, and text reports; adopts a pre-trained cross-modal fusion neural network model to extract high-dimensional visual features, spectral features, and text semantic features corresponding to various types of data, and then calculates the correlation weights between different modalities by constructing a cross-modal attention mechanism module, filters strongly correlated features, and generates a global correlation feature vector of the site selection environment; inputs the global correlation feature vector into a multi-task decision head, calculates the comprehensive score of each candidate site through a weighted fusion strategy, and outputs the final site selection decision result. First, the multi-source heterogeneous raw data involved in substation site selection are uniformly preprocessed and standardized to construct a multimodal dataset. Then, independent feature extraction of each modality is completed through multiple types of networks. Relying on a cross-modal attention mechanism, deep interactive fusion and effective filtering of multimodal features are achieved. Finally, the site selection result is output by combining a weighted score from a multi-task decision head. This breaks through the limitation of existing technologies that can only use single-modal data for site selection analysis, and realizes the full fusion and in-depth mining of multi-source heterogeneous data such as remote sensing images, infrared thermal imaging, acoustic signals, and text reports. It solves the problems of existing AI site selection decision systems being unable to fully integrate multi-source heterogeneous data, resulting in incomplete site selection environment perception analysis, low analysis accuracy, poor decision reliability, and deviations between site selection results and actual needs.

[0066] Example 3 Please see Figure 2 The figure shows a substation site selection system based on artificial intelligence proposed in the third embodiment of the present invention. The system includes: The acquisition module 100 is used to acquire multi-source heterogeneous raw data of the area to be selected, and to perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions. The generation module 200 is used to perform feature extraction and interaction analysis on a multimodal location dataset using a pre-trained cross-modal fusion neural network model, generating a globally correlated feature vector of the location environment, specifically including: High-dimensional visual features of remote sensing images and infrared thermal images are extracted using convolutional neural networks or visual Transformers, spectral features of voiceprint signals are extracted using audio neural networks, and textual semantic features of text reports are extracted using pre-trained language models. A cross-modal attention mechanism module is constructed, using text semantic features as query vectors and visual and spectral features as key vectors to calculate the association weights between different modalities, thereby filtering out visual and auditory features that are strongly related to the location task and obtaining the fused global association feature vector. The location selection module 300 is used to input the global correlation feature vector into the multi-task decision head, calculate the comprehensive score of each candidate point through a weighted fusion strategy, and output the final location selection decision result.

[0067] The functions or operation steps implemented by the above modules are largely the same as those in the above method embodiments, and will not be repeated here.

[0068] Example 4 In another aspect, the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in any one of Embodiments 1 to 2 above.

[0069] Example 5 In another aspect, the present invention provides an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of any one of the methods described in Embodiments 1 to 2 above.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0072] More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0073] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0074] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0075] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A substation site selection method based on artificial intelligence, characterized in that, The method includes: Acquire multi-source heterogeneous raw data of the proposed site area, and perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions. A pre-trained cross-modal fusion neural network model is used to extract features and perform interactive analysis on a multimodal location dataset, generating a globally correlated feature vector of the location environment, specifically including: High-dimensional visual features of remote sensing images and infrared thermal images are extracted using convolutional neural networks or visual Transformers, spectral features of voiceprint signals are extracted using audio neural networks, and textual semantic features of text reports are extracted using pre-trained language models. A cross-modal attention mechanism module is constructed, using text semantic features as query vectors and visual and spectral features as key vectors to calculate the association weights between different modalities, thereby filtering out visual and auditory features that are strongly related to the location task and obtaining the fused global association feature vector. The global correlation feature vector is input into the multi-task decision head, and the comprehensive score of each candidate point is calculated through a weighted fusion strategy, and the final site selection decision result is output.

2. The substation site selection method based on artificial intelligence according to claim 1, characterized in that, The steps for spatiotemporal reference alignment and preprocessing of multi-source heterogeneous raw data include: The remote sensing image and the infrared thermal image are registered at the pixel level using geographic information system coordinates, and the voiceprint signal is mapped to the corresponding spatial coordinate point; Key geographic entities and constraints in text reports are extracted using natural language processing techniques and vectorized into text semantic features.

3. The substation site selection method based on artificial intelligence according to claim 2, characterized in that, The step of extracting key geographic entities and constraints from the text report using natural language processing technology and vectorizing them into text semantic features includes: Construct a location selection knowledge graph based on a graph structure. The nodes of the knowledge graph include geographic entities, geological parameters and planning constraints, and the edges represent the spatial adjacency or logical constraint relationships between nodes. A pre-trained graph attention network is used to reason about the location knowledge graph, calculate the importance score of each node relative to the location target, and select key constraint nodes based on the importance score. The text fragments corresponding to the key constraint nodes are input into a bidirectional long short-term memory network to extract context semantic vectors. The context semantic vectors are then concatenated with the topology vectors output by the graph attention network to obtain the final text semantic features.

4. The substation site selection method based on artificial intelligence according to claim 1, characterized in that, The steps of constructing the cross-modal attention mechanism module, which uses text semantic features as query vectors and visual and spectral features as key vectors, to calculate the association weights between different modalities, thereby filtering out visual and auditory features strongly related to the location selection task, and obtaining the fused global association feature vector, include: Text semantic features are mapped to query vectors through a linear projection layer. The high-dimensional visual features of remote sensing images and infrared thermal images, as well as the spectral features of the voiceprint signal, are mapped into visual key vectors through independent linear projection layers. Visual value vector Auditory key vector and auditory value vector ; Calculate query vector With visual key vector The dot product similarity is used to obtain the text-visual attention map, and the query vector is calculated simultaneously. With auditory key vector The dot product similarity is used to obtain the text-auditory attention map; Normalize the text-visual attention map and the text-auditory attention map respectively to obtain the visual association weight matrix. Hearing association weight matrix ; With visual value vector As a query, with auditory key vector As the key, with auditory value vector As values, the guiding weights of visual features on auditory features are calculated to generate visually guided auditory enhancement features. ; Visual association weight matrix Acting on the visual value vector Weighted visual features are obtained The auditory association weight matrix Auditory enhancement features acting on visual guidance Weighted auditory features were obtained. ; Weighted visual features and weighted auditory features Channel-level concatenation is performed, and feature fusion is carried out through a multilayer perceptron to output a globally correlated feature vector.

5. The substation site selection method based on artificial intelligence according to claim 1, characterized in that, The steps of inputting the global correlation feature vector into the multi-task decision head, calculating the comprehensive score of each candidate point through a weighted fusion strategy, and outputting the final site selection decision result include: The multi-task decision head includes a suitability classifier and a risk regressor, which output the suitability probability and potential risk value of the site selection, respectively. During the decision-making process, a conflict resolution algorithm based on evidence theory is introduced. When there is a logical conflict between the abnormal area shown by infrared thermal imaging and the safety description in the text report, the decision weight of the infrared mode is automatically increased, and a secondary verification is performed by combining the transient abnormal features in the voiceprint signal. Finally, a site selection report containing recommended coordinates, risk level and decision basis is generated.

6. The substation site selection method based on artificial intelligence according to claim 5, characterized in that, The introduced conflict resolution algorithm based on evidence theory, when there is a logical conflict between the abnormal area displayed by infrared thermal imaging and the security description in the text report, automatically increases the decision weight of the infrared mode, including the following steps: The thermal anomaly features extracted from the infrared thermal image and the security constraint semantic features extracted from the text report are respectively regarded as two independent evidence bodies, with the infrared thermal image evidence body being the most important. Corresponding to the confidence level of thermal anomaly evidence, the text report evidence body Corresponding to the security description confidence evidence, both evidence bodies are mapped to a pre-defined location security identification framework. The identification framework It includes three identification elements: "safety", "risk", and "uncertainty"; Calculate the two pieces of evidence , The basic probability assignment function, where The BPA function is calculated based on the temperature difference, area ratio, and duration of continuous occurrence of abnormal regions in the infrared thermal image. The BPA function is calculated based on the semantic similarity, frequency of occurrence of security constraint keywords in the text report, and importance scores of constraint nodes in the knowledge graph. The Dempster-Shafer synthesis rule is used to calculate the fusion probability of two pieces of evidence, which is then quantified by introducing a conflict coefficient. and The degree of logical conflict is determined. When the conflict coefficient is greater than or equal to the preset conflict threshold, it is determined that there is a logical conflict between the abnormal area displayed by the infrared thermal imaging and the security description in the text report. Using the conflict coefficient as input, and combining the hazard level weight of thermal anomaly risk in the substation site selection scenario, the weight adjustment increment of the infrared mode is calculated through a nonlinear mapping function, which is trained based on the review data of historical site selection conflict cases. Based on the adjusted infrared modal weights, the input feature weights of the suitability classifier and risk regressor in the multi-task decision head are recalculated. The adjusted weights are then applied to the high-dimensional visual features of infrared thermal imaging, while the weight ratio of the semantic features of the text report is simultaneously reduced, thus completing the feature weight calibration after conflict resolution.

7. The substation site selection method based on artificial intelligence according to claim 5, characterized in that, The step of performing secondary verification by combining transient anomaly features in the voiceprint signal, and finally generating a site selection report containing recommended coordinates, risk level, and decision-making basis includes: Transient anomaly features within a preset frequency band are extracted from the voiceprint signal. An algorithm combining wavelet packet decomposition and short-time Fourier transform is used to decompose the voiceprint signal at multiple scales. Transient signal segments with a duration of a preset time and an amplitude exceeding a preset multiple of the normal voiceprint baseline are selected as candidate sets of transient anomaly features. The candidate set of transient anomaly features is input into a pre-trained audio classification network. Combined with a sample library of typical interference sound patterns in the substation site selection scenario, the candidate set is classified and identified, and invalid interference transient features are eliminated to obtain effective transient anomaly features related to thermal anomalies. Calculate the spatial correlation and feature similarity between effective transient anomaly features and infrared thermal anomaly features. If the overlap between the spatial coordinates corresponding to the effective transient anomaly features and the infrared thermal anomaly region is greater than a threshold, and the feature similarity is greater than a threshold, then the effectiveness of the infrared thermal anomaly features is confirmed, and the current site selection risk assessment results are strengthened.

8. A substation site selection system based on artificial intelligence, characterized in that, The system includes: The acquisition module is used to acquire multi-source heterogeneous raw data of the site selection area, and to perform spatiotemporal benchmark alignment and preprocessing on the multi-source heterogeneous raw data to construct a multimodal site selection dataset. The multi-source heterogeneous raw data includes at least remote sensing images reflecting geospatial characteristics, infrared thermal images reflecting the distribution of equipment or surface thermal anomalies, acoustic fingerprint signals reflecting environmental sound field characteristics, and text reports containing geological survey or planning restrictions. The generation module is used to perform feature extraction and interaction analysis on multimodal location datasets using a pre-trained cross-modal fusion neural network model, generating a globally correlated feature vector of the location environment, specifically including: High-dimensional visual features of remote sensing images and infrared thermal images are extracted using convolutional neural networks or visual Transformers, spectral features of voiceprint signals are extracted using audio neural networks, and textual semantic features of text reports are extracted using pre-trained language models. A cross-modal attention mechanism module is constructed, using text semantic features as query vectors and visual and spectral features as key vectors to calculate the association weights between different modalities, thereby filtering out visual and auditory features that are strongly related to the location task and obtaining the fused global association feature vector. The location selection module is used to input the global correlation feature vector into the multi-task decision head, calculate the comprehensive score of each candidate point through a weighted fusion strategy, and output the final location selection decision result.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 7.

Images