1597 results about "Time sequence" patented technology

The invention relates to the technical field of marine environment prediction, in particular to a marine environment multi-modal fusion prediction method and system based on digital twinning. The method comprises the following steps: capturing time-space correlation characteristics of multi-modal data of a marine environment based on a dynamic multi-modal graph neural network; performing multi-scale feature fusion on the space-time correlation features by using a multi-scale gating unit to obtain comprehensive feature representation; predicting the comprehensive feature representation by using a mixed time sequence prediction framework to obtain preliminary marine environment prediction data, including short-term dynamic modeling and long-term trend modeling; and performing noise fitting on the preliminary marine environment prediction data by using a generative adversarial network to generate final marine environment prediction data. According to the invention, through innovative combination of dynamic multi-modal data fusion, a mixed time sequence prediction framework and a multi-scale gating Tanh unit, the problems of insufficient data fusion, limited prediction precision and poor dynamic response capability in the traditional marine environment monitoring and prediction technology are effectively solved.

The invention relates to the technical field of electric power short-term load prediction, in particular to an industrial park short-term load prediction method and system based on a Diffusion-Transform-AR model. Acquiring non-stationary historical load data and meteorological data of the industrial park, and performing missing value processing on the data; performing correlation analysis on the data by using the maximum information coefficient, selecting meteorological data with relatively high correlation, and reprocessing time data; performing time sequence decomposition on the load data by using a seasonal trend decomposition method to obtain trend term, seasonal term and residual term features; constructing a diffusion Transform prediction model, and taking the residual term, the meteorological data and the time data as input prediction residual; constructing an autoregression prediction model and taking the trend term as an input prediction trend; according to the method, the advantages of various models are combined, time sequence features are deeply mined, the prediction precision and the model generalization ability are effectively improved, and the method is suitable for complex industrial park load prediction scenes.

The invention provides a power prediction precision improvement method and system for a wind and light station, and relates to the technical field of new energy power generation, and the method comprises the steps: obtaining the historical power, weather and equipment operation data of the wind and light station through a data preprocessing module, and carrying out the time sequence decomposition, spatial interpolation, abnormal value detection and correction, and obtaining the fusion feature data. And constructing a deep learning prediction model based on the fused feature data, extracting spatial features by using a convolutional neural network, processing the spatial features by using a long-short-term memory network to obtain an initial prediction result, introducing an attention mechanism to perform weight allocation to obtain weighted features, and combining to obtain power prediction data. And judging the weather type according to the real-time meteorological condition, and correcting by adopting the matched correction model when the extreme weather is identified. And finally, an error compensation model is constructed based on historical prediction errors for compensation, a final power prediction result is obtained, error correction is performed according to actual power data, and compensation model parameters are updated.

The invention provides a marine environment time sequence prediction method and system based on a recurrent neural network, and relates to the technical field of time sequence prediction, and the method comprises the steps: obtaining marine environment historical observation data, carrying out the preprocessing of the marine environment historical observation data, and removing abnormal values and missing values from the marine environment historical observation data; according to the preprocessed marine environment historical observation data, a recurrent neural network model is constructed, and the recurrent neural network model comprises an input layer, two stacked long short-term memory (LSTM) network layers and an output layer; acquiring marine environment observation data at the current moment, and inputting the marine environment observation data into the trained recurrent neural network model; predicting marine environment data of a plurality of time steps in the future through a recurrent neural network model, and obtaining a marine environment prediction result in a preset time; and according to the marine environment prediction result, identifying a hot spot region of marine environment change, and generating a visual result representing a dynamic change rule of the marine environment in the prediction time period.

The invention relates to the technical field of data processing, and discloses a time sequence feature analysis method and system based on multi-dimensional data. The method comprises the following steps: performing spatial-temporal feature extraction and dimension reduction processing on a multi-dimensional time sequence precipitation data set to obtain a precipitation spatial-temporal feature vector set; performing multi-scale spatio-temporal pattern recognition and clustering analysis on the precipitation spatio-temporal feature vector set to obtain a precipitation spatio-temporal pattern library, a clustering result and an abnormal precipitation pattern list; constructing a deep learning rainfall prediction model to obtain an end-to-end global atmospheric reanalysis rainfall fusion prediction model; performing dynamic updating and online learning on the global atmospheric reanalysis rainfall fusion prediction model to obtain a target rainfall space-time prediction model; and performing multi-dimensional space-time analysis on the target rainfall space-time prediction model to obtain a global atmosphere reanalysis rainfall fusion data analysis report and a weather processing assembly line scheduling strategy. According to the invention, the accuracy of time sequence feature analysis based on multi-dimensional data is improved.

The invention discloses a submarine cable risk dynamic assessment method and system based on multi-modal deep learning, and belongs to the field of marine infrastructure operation and maintenance. Aiming at the problems of incomplete data coverage, unreal generated scene, low evaluation reliability and the like in the prior art, the method comprises the following steps of: 1) constructing a multi-source heterogeneous data set containing six types of data including geology, ocean, ships, biology and the like, and realizing data alignment by adopting space-time grid coding; 2) designing a physical constraint generative adversarial network, and generating risk scene data conforming to a fluid mechanics law through a Navier-Stokes equation constraint; 3) creating a hierarchical space-time fusion network (HST-Transform), and combining CNN spatial feature extraction, a time sequence attention mechanism and a dynamic memory module to realize multi-modal fusion; according to the method, the detection rate of rare risk events is increased by 62%, the evaluation accuracy rate reaches 91.7%, the false alarm rate is reduced by 34% compared with a traditional method, and submarine cable breakage accidents can be effectively prevented.

The invention discloses a power distribution network power supply reliability evaluation method and system based on data analysis, and relates to the technical field of power distribution network detection, a meteorological station, a GIS and an equipment monitoring terminal collect temperature, rainfall, wind speed, load and fault record data, and after time sequence alignment and interpolation correction, a unified environment data set is generated to provide an accurate basis for model construction; secondly, using historical and online data, calling a nonlinear mapping function and an equipment function to train a meteorological-fault rate model, outputting the equipment fault probability in real time, and then calculating a system risk index and a line risk value in combination with confidence-weighted weather forecast and equipment load coupling operation to accurately identify a high-risk area; priority scheduling, load transfer and graded load shedding are carried out on the high-risk area, fault information is transmitted back, closed-loop control is achieved, and the operation efficiency is improved.

The invention discloses a solar radiation space-time prediction method and system based on physical information constraint and a neural network, and the method comprises the steps: collecting multi-dimensional time sequence meteorological data, extracting high-dimensional time sequence dynamic features, converting geographic space data into a fuzzy set, and carrying out the defuzzification of the fuzzy set through an inference rule, thereby obtaining geographic space features; and a gating mechanism is adopted to realize deep fusion of the space-time features to generate high-dimensional space-time fusion features. In a model training stage, an energy conservation equation is introduced into an optimization process, a physical residual error is constructed by calculating a time derivative and a space derivative of a predicted value, a physical constraint total loss function is formed in combination with a data loss item, and model parameters are updated by using a gradient descent method. According to the method, the accuracy and reliability of a prediction result are remarkably improved while the calculation efficiency is ensured, and the method is particularly suitable for solar radiation prediction under complex meteorological conditions; according to the method, abnormal prediction caused by data noise can be effectively corrected, and a solution with physical rationality and data adaptability is provided for the fields of solar resource evaluation, photovoltaic power generation power prediction and the like.

The invention relates to the technical field of communication, in particular to a low-delay data transmission system and method for marine observation data, and the system comprises a data collection module which collects the multi-modal data of a marine environment, and carries out the priority classification of the data according to the real-time performance and importance; the route optimization module dynamically adjusts a multi-path fragmentation strategy and single-path task distribution in combination with a reinforcement learning algorithm and path optimization data; the communication module supports multi-mode communication modes such as laser communication and low-orbit satellite communication, and monitors the link state in real time; and the data fusion and analysis module performs dimension reduction, fusion and real-time analysis on the data by using manifold learning and a graph neural network, and meanwhile, predicts a long-term trend through time sequence analysis to generate observation report data. The method effectively reduces the transmission delay of high-priority data, improves the response efficiency of the system to abnormal events, and is suitable for the scenes of marine disaster early warning, environment monitoring and the like.

The invention provides a mine abnormal event real-time identification method and system based on time sequence characteristics, and relates to the technical field of mode identification, and the method comprises the steps: carrying out the time-space alignment and semantic annotation of multi-modal monitoring data, and constructing a time sequence knowledge graph; calculating a dynamic association weight between entities, and analyzing a risk propagation path; predicting a risk situation based on a historical evolution rule; and dynamically generating a differential early warning strategy and establishing a closed-loop tracking system. According to the invention, early identification, accurate prediction and efficient disposal of mine safety risks can be realized, and the mine safety management level is improved.

The invention discloses a multi-source multi-temporal-spatial feature fusion short and temporary rainfall prediction method (M4Cester), which realizes layered temporal-spatial feature extraction through a multi-space multi-time aggregator (MMA), and comprises a multi-scale patch embedding (MSPE) module, a cross-scale perception optimization (CPR) module and a multi-time self-attention (MTS) module. The first and second spatial features are respectively used for capturing local-global spatial features, dynamically balancing cross-scale information and mining a time sequence dependency relationship; meanwhile, a bidirectional bridging fusion module (MSFM) is designed, bidirectional alignment and enhancement of radar and satellite features are realized by using a cross attention mechanism, and modal difference is relieved through residual connection. Experiments on a weather data set in the Yangtze River Delta region show that the key success index (CSI) reaches 0.267 and the HSS reaches 0.372 in heavy rainfall (greater than or equal to 50dBZ) prediction by the method, which are obviously improved compared with the existing advanced model, particularly, the limitation of single-source data is effectively overcome in the prediction of a convection initiation (CI) event, and a high-precision solution is provided for short and temporary rainfall prediction.

The invention discloses a power distribution network overhead line risk refined early warning method and system. The method comprises the following steps: collecting meteorological time sequence data and power distribution network topology and line parameter data; and preprocessing the data, and constructing a prediction data set. And establishing a fault probability prediction model based on the meteorological data and the line parameters. Calculating the fault probability by using the mathematical model, correcting the model, training in combination with static topology and dynamic meteorological data, and constructing a fault probability network model. And analyzing the importance of the nodes and lines of the power distribution network, and evaluating the load loss proportion of the nodes. And carrying out normalization processing on the importance indexes. And comprehensively calculating the predicted fault probability, the node loss load and the importance index to obtain a branch risk parameter. Fine early warning is carried out according to a set early warning threshold value, and dynamic adjustment is carried out to adapt to the actual operation condition. The method has the advantages that comprehensive risk assessment is provided, and the accuracy of fault prediction and the timeliness of risk early warning are improved.

The invention discloses a power system false data injection attack detection method and system. Measurement data of a power distribution system are acquired; constructing an attack sequence, and injecting the constructed attack sequence into the obtained measurement data to simulate a false data attack; constructing a false data injection attack detection model based on the expansion causal convolution and the recurrent neural network to analyze the state change of the power distribution system; and judging an attack behavior according to the analyzed state of the power distribution system. According to the method, time sequence data can be processed, the long-term dependency relationship and the complex mode can be captured, and the response capability to the complex attack mode can be improved, so that the false data injection attack can be effectively detected and prevented, and the safety and reliability of the power distribution system can be maintained.

The invention provides a water conservancy data acquisition supervision method and system based on big data analysis. The method comprises the following steps: firstly, obtaining dam body surface temperature field data, reservoir water level time sequence monitoring data and a dam body three-dimensional structure, and carrying out environmental radiation interference elimination processing on the temperature field data; secondly, extracting temperature fluctuation amplitude time sequence characteristics and water level change rate in the corrected temperature field data, and constructing a matrix reflecting correlation strength of the temperature fluctuation amplitude time sequence characteristics and the water level change rate through dynamic correlation analysis; and identifying temperature fluctuation abnormal points by using the matrix, and generating a seepage correlation map in combination with the water level change rate. And dividing seepage state categories according to the water level change rate, and generating a leakage risk probability distribution diagram by combining periodic trend prediction. And finally, fusing the graph with a dam body three-dimensional structure to generate positioning supervision information with a seepage abnormal identifier. The technical scheme provided by the invention can improve the efficiency and accuracy of water conservancy data acquisition supervision.

The invention relates to a power equipment fault diagnosis method based on time sequence-spectrogram-discrete data multi-modal fusion, and belongs to the field of power equipment fault diagnosis. A multi-modal feature fusion technology based on an attention mechanism is adopted, and a power equipment fault diagnosis model based on multi-modal information fusion is constructed. Compared with a traditional deep learning fault diagnosis method, the problem that a fault single-modal sample is insufficient in diagnosis model accuracy can be solved through multi-modal data fusion, and the universality and diagnosis precision of the model are improved.

The invention provides a power load long time sequence prediction method and a related device, and relates to the technical field of power load prediction. The method comprises the steps of collecting power load time sequence data; preprocessing the power load time sequence data; decomposing the preprocessed power load time sequence data into a seasonal component and a trend component through a mobile decomposition method; inputting the trend components into a trend clustering enhancement network, performing channel clustering on the trend components through a K-means clustering algorithm, applying independent linear transformation to each clustering group, and extracting long-term trend features; inputting the seasonal components into a seasonal causal attention network, and extracting short-term seasonal features through a causal mask mechanism and a multi-head attention mechanism; fusing the long-term trend features and the short-term seasonal features to generate fused features; and inputting the fusion features into a decoder, and outputting a power load prediction result through a full connection layer. The change trend of the power load data is fully considered, and the prediction precision of the power load long sequence is improved.

The invention relates to the technical field of thermal power plants, in particular to a thermal power plant load prediction method and system. The method comprises the following steps: acquiring real-time operation data of a thermal power plant, wherein the real-time operation data comprises historical load data, unit operation parameter data, environment parameter data and power grid dispatching data; performing data missing processing based on a time sequence interpolation method on the real-time operation data, and detecting and correcting an abnormal value to obtain operation preprocessing data; multi-scale time feature extraction is carried out on the operation preprocessing data to obtain time feature data, and multi-scale time features comprise hour features, day features and weekly features; and constructing a multi-level load prediction model based on the time characteristic data. According to the method, through multi-scale time feature extraction, load features of different time levels such as hours, days and weeks are comprehensively covered, and richer input information is provided for prediction.

The invention relates to the technical field of intelligent electronic data processing in an underground comprehensive pipe gallery, and discloses an underground comprehensive pipe gallery intelligent data fusion and dynamic early warning system and method, and the system comprises a data collection module which is used for receiving monitoring data containing timestamps from a plurality of sensors in the underground comprehensive pipe gallery in real time; the time sequence causal chain synchronization module dynamically constructs a causal time sequence chain between the data according to a predefined physical topological structure; the lightweight causal pruning device module predefines a causal trunk path based on physical topology and activates complete causal chain construction when the causal trunk path is abnormal; and the dynamic early warning module is used for sending out early warning information based on a causal time sequence chain analysis result. According to the method, through micro timestamp deviation calibration and dynamic causal chain construction based on physical topology, real causes of abnormal events can be accurately identified, early symptoms omissed due to neglecting of time sequence relevance of a traditional system are captured, and precious time is won for timely disposal.

The invention relates to the field of energy management, and discloses an industrial park load prediction method based on artificial intelligence, and the method comprises the steps: collecting historical load, production plan, weather and holiday and festival information through multi-source data; through data preprocessing, a time sequence feature extraction module fusing an attention mechanism and LSTM and a deep learning model integrating a sudden change adaptation module are constructed, and high-precision load prediction is realized. Wherein the abrupt change adaptation module dynamically adjusts model parameters to cope with load abrupt change through sliding window statistical feature monitoring, incremental learning and GAN simulation abrupt change scenes; and a real-time feedback mechanism further optimizes the prediction result, and generates a power dispatching suggestion in combination with a dynamic electricity price strategy. According to the method, the problems of large prediction deviation and poor adaptability of a traditional model in a load sudden change scene are solved, and the efficiency and stability of industrial park energy management are remarkably improved.

The invention provides a feeder terminal fault detection method, system and device, a medium and a program product, and the method comprises the steps: obtaining multi-source heterogeneous data comprising feeder terminal operation data, a topological graph structure of a power distribution network and external sensing data, the multi-source heterogeneous data comprises at least one kind of structured or unstructured time sequence data, and the external sensing data comprises at least one kind of structured or unstructured time sequence data; the external sensing data comprises meteorological data, geographic space information and historical fault records; and preprocessing the multi-source heterogeneous data, inputting the preprocessed time-aligned multi-source heterogeneous data into the trained time-space diagram neural network model to extract features and perform joint modeling, and outputting a fault type classification result and a fault probability distribution result corresponding to each feeder terminal node in the power distribution network. According to the method, a multi-source heterogeneous data fusion and time-space diagram neural network modeling mechanism is introduced, so that the model has a dynamic response capability to complex environment changes, and the identification precision of a potential fault mode is effectively enhanced.

The invention provides a deep learning traffic flow prediction method based on meteorological information fusion, and the method comprises the following steps: 1, carrying out the preprocessing of traffic data and weather data; step 2, establishing a time alignment module to solve the time dislocation problem of traffic data and weather data; step 3, carrying out embedded coding on the traffic data and the weather data; step 4, establishing an enhanced space-time convolutional network, extracting space-time features from the traffic data through the enhanced space-time convolutional network, and extracting time sequence features from the meteorological data through the enhanced space-time convolutional network; 5, establishing a comparative learning module, and performing feature fusion and comparative learning on the spatio-temporal features and the time sequence features; and step 6, generating a predicted value of the traffic flow. According to the method, the defects of an existing traffic flow prediction method in the aspects of multi-source data integration, long-time dependence capture, feature fusion and the like are effectively overcome, and a new technical approach is provided for improving the precision of traffic flow prediction and the universality of practical application.

The invention discloses an urban tunnel intelligent safety operation and maintenance method and system combined with multiple data, particularly relates to the field of intelligent traffic operation and maintenance, and is used for solving the problems of prediction deviation and congestion management caused by tunnel flow fluctuation. In combination with traffic parameters of a multi-point detector, a wave similarity analysis model is utilized to accurately evaluate a propagation path and a target section of fluctuation; meanwhile, by means of a time sequence superposition correction mechanism, a historical fluctuation rule and real-time data are fused, a traffic flow transmission prediction curve after dynamic correction is generated, and balance between physical constraint and prediction precision is achieved; finally, key indexes are output through dynamic updating of the flow prediction model, efficient linkage of signal coordination and induction management and control modules is guided, congestion accumulation and traffic risks caused by fluctuation propagation can be reduced to the maximum extent, the tunnel operation and maintenance efficiency is optimized, and the traffic safety guarantee capacity is enhanced.

The invention provides a wind power prediction method and system based on time sequence decomposition and multi-model fusion, and the method comprises the steps: collecting the historical power generation power and meteorological data of a target wind power plant; decomposing the historical power generation power and the meteorological data to obtain a trend component, a seasonal component and a residual component; fusing with meteorological data to construct a trend feature matrix, a periodic feature matrix and a residual feature matrix; different modeling schemes are adopted to construct corresponding single models; dividing into a training set, a verification set and a test set according to a time sequence; performing training optimization on the single model by using the training set, the verification set and the test set, and constructing a wind power short-term power prediction model; and inputting the real-time meteorological data and the generated power to the wind power short-term power prediction model, and outputting the generated power prediction value of the target wind power plant, thereby effectively improving the comprehensiveness, accuracy and stability of model prediction.

The invention relates to the technical field of photovoltaic power prediction, and discloses a regional distributed photovoltaic power prediction method and system, and the photovoltaic power prediction comprises the steps: S1, obtaining meteorological data in real time, introducing a time sequence processing capability model, and capturing the nonlinear influence of the weather change on the photovoltaic power; s2, introducing an online learning and self-adaptive mechanism, enabling the prediction model to have an online learning capability, and dynamically adjusting model parameters along with the inflow of new real-time meteorological data; s3, implementing rolling prediction based on a short time window, and dynamically adjusting the power generation plan: when a sudden weather change is detected, triggering a quick response program, and adjusting the power generation plan according to the latest prediction result; and S4, integrating model prediction results, and adopting a probability prediction method to provide probability distribution of photovoltaic power output instead of a deterministic numerical value, so as to reflect an uncertainty range of future photovoltaic power change.

The invention is suitable for the technical field of health monitoring, and provides a health monitoring method and system based on time sequence Transform and data fusion, and the method comprises the following steps: collecting multi-modal biosensor data; carrying out asynchronous time processing and normalization processing on the acquired multi-mode biosensor data; extracting features from the processed multi-modal biosensor data by using TimesNet, forming an independent input channel by the data of each modal, performing feature extraction on the data of each modal, extracting time sequence features of different scales and abstract levels, and performing feature extraction on the extracted time sequence features of different scales and abstract levels; the extracted features are input into a time sequence Transform network for feature fusion, time sequence prediction and anomaly monitoring; carrying out online platform display on the obtained prediction data, and giving an alarm when abnormity exists in the prediction data; through the time sequence Transform model, the health state of the worker can be monitored in real time, potential abnormal behaviors can be found and diagnosed in time, and the safety management level is improved.

The invention discloses a marine water environment abnormal state secondary and derivation probability modeling and estimating method, which belongs to the field of state modeling estimation, and comprises the following steps: acquiring real-time data flow of marine environment parameters, performing segmentation processing by adopting a sliding window mechanism, and reducing data dimension by utilizing principal component analysis. Then, constructing a graph structure model based on a time sequence, and capturing correlation and a space-time dependency relationship among parameters; and performing nonlinear mapping on the graph structure model by adopting a radial basis function core, and capturing complex nonlinear characteristics among the marine environment parameters. And finally, classifying the abnormal states by using a multi-class support vector machine algorithm, and outputting secondary and derivative probability estimation results of the abnormal states in real time. According to the method, abnormal events in the marine environment can be found in time, dynamic modeling support can be provided for marine environment monitoring, and the accuracy and real-time performance of marine environment monitoring are remarkably improved.

The invention discloses an industry return on investment real-time acquisition method based on multi-source heterogeneous data, and relates to the technical field of data analysis and processing, and the method comprises the following steps: normalizing a multi-source timestamp and retaining metadata information based on a UTC reference, an adjustable gain and a power exponent; in the streaming processing, reordering of out-of-order data and marking of expired data are realized by setting a buffer queue and a watermark threshold value; when null value fields or cross-source conflicts are detected, interpolation and conflict processing are carried out, and complete data subjected to consistency correction are output; then, an external high-precision reference or a cross-correlation function is used for further fine tuning the timestamp at a millisecond level, and if the adjustment amplitude exceeds a safety boundary, the timestamp is marked as suspicious; and finally, the corrected data is distributed to multiple nodes according to a fragment mapping strategy, and fault-tolerant consensus and difference repair are triggered under fault duration judgment, so that high-precision time sequence alignment and high availability in a massive concurrent scene are kept, data loss or precision attenuation caused by time sequence inconsistency and node faults is avoided, and the method can be widely applied to high-frequency analysis.

The invention discloses a photovoltaic power generation power prediction method and system based on a large language model. The method comprises the following steps: converting historical power data and numerical weather forecast data into time sequence embedded representation; through cross-modal semantic alignment, semantic embedding representation is generated; constructing a natural language prompt containing task context information, encoding the natural language prompt into prompt embedding, combining prompt embedding with semantic embedding representation to form a fusion input sequence, inputting the fusion input sequence into a pre-trained large language model, and outputting implicit features; synchronously generating an initial power prediction result and a weather prediction result obtained by correcting the numerical weather prediction data through a parallel collaborative prediction mechanism; and taking the meteorological prediction result as a correction signal, performing joint optimization on the preliminary power prediction result, and outputting a power generation power prediction value. According to the method, the problem of deep fusion of heterogeneous data is effectively solved, and the prediction accuracy is improved.

The invention discloses an energy storage system state evolution trend prediction method based on multi-source data fusion. The method comprises the steps of terminal voltage, current and temperature time sequence data acquisition, time sequence segmentation normalization, multi-physics field coupling feature construction, trend prediction model construction and training and energy storage system state evolution trend prediction. According to the method, the distinguishing capacity of the model for charging and discharging physical characteristics is improved, meanwhile, the voltage change rate, the multi-dimensional feature vector of the differential internal resistance and the thermal-electric coupling effect and the explicit encoding electric-thermal-resistance coupling relation are constructed, the transient response and the temperature hysteresis effect can be effectively captured, and then the model can be used for analyzing the charging and discharging physical characteristics. A degradation-aware cross-cycle feature extraction and gating mechanism is adopted, short-term fluctuation and long-term trend are adaptively balanced in multi-scale prediction, the prediction conflict problem is relieved, finally, physical constraints based on the electrochemical law and the internal resistance temperature characteristic are embedded in a loss function, it is ensured that the prediction result is accurate in numerical value and conforms to the physical law, and the prediction accuracy is improved. And generation of physically impossible solutions is avoided.

The invention provides an ocean subsurface thermohaline reconstruction method based on multi-scale spatial-temporal feature fusion, and solves the technical problem of poor thermohaline reconstruction precision caused by incapability of capturing deep-level spatial-temporal dependence in ocean data in the prior art. The method comprises the steps of obtaining an ocean temperature-salinity anomaly feature data set, performing preprocessing to obtain time sequence data and space sequence data, inputting a multi-scale spatial-temporal feature fusion network model to perform feature fusion to obtain a temperature-salinity anomaly value, performing reverse normalization, and superposing a climate average value to obtain a temperature-salinity reconstruction result; the multi-scale spatio-temporal feature fusion network model comprises a spatial feature extraction branch, a temporal feature extraction branch and a spatio-temporal feature fusion module. The method can be widely applied to the technical field of marine science.