CN-121997031-A - Double-branch two-stage sea surface temperature forecasting method based on multi-element input

CN121997031ACN 121997031 ACN121997031 ACN 121997031ACN-121997031-A

Abstract

The invention relates to the technical field of marine environment forecasting, and discloses a double-branch two-stage marine surface temperature forecasting method based on multi-element input. The method comprises the steps of obtaining and preprocessing multi-element data such as sea surface temperature, 2m temperature and atmospheric top incident solar radiation, constructing a double-branch collaborative optimization deep learning model, training the model, inputting the data for a plurality of days before a time to be forecasted into the model, and generating a future sea surface temperature forecast result. The short-term prediction branch extracts space-time characteristics through ConvGRU and multi-scale convolution to predict short-term results, the middle-long term prediction branch models long-range dependence through adaptive weighting and a transducer encoder to predict middle-long term results, and the future multi-day prediction is generated through autoregressive rolling. According to the invention, through the cooperation of the double branches, error accumulation is inhibited, the accuracy and stability of long-term prediction in sea surface temperature are obviously improved, and efficient and accurate technical support can be provided for ocean resource development.

Inventors

SUN WEIFU
ZHANG YUHAO
ZHOU ZHIXUAN
LI XINFANG
PENG XINRAN

Assignees

自然资源部第一海洋研究所

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (10)

1. A multi-element input-based double-branch two-stage sea surface temperature forecasting method is characterized by comprising the following steps: Step S1, acquiring ocean-atmosphere multi-element re-analysis data of a target sea area, and preprocessing the data to construct a training data set, wherein the multi-element re-analysis data comprise sea surface temperature, 2 meters temperature and atmospheric top incident solar radiation; Step S2, constructing a double-branch collaborative optimization deep learning model, wherein the model comprises a short-term prediction branch and a medium-long term prediction branch, the short-term prediction branch is used for extracting space-time characteristics from input sea surface temperature data and generating a sea surface temperature prediction result of a first period, and the medium-long term prediction branch is used for generating a sea surface temperature prediction result of a second period based on self-adaptive weighted multi-source variables and long-distance space-time dependency modeling, wherein the first period is adjacent to the second period and the second period is longer; S3, training the model by using the training data set; And S4, inputting multi-element analysis data of a plurality of continuous days before the time to be forecasted into a trained model, generating a sea surface temperature forecasting result of a first period in the future through the short-term forecasting branch, and generating a sea surface temperature forecasting result of a second period in the future through the medium-term forecasting branch.
2. The method for forecasting the sea surface temperature in two stages based on multi-element input is characterized in that the short-term forecasting branch is formed by sequentially connecting a front ConvGRU, a multi-scale space information fusion module, a characteristic splicing fusion unit and a rear ConvGRU; the preamble ConvGRU is used for extracting a preliminary time sequence feature; the multi-scale space information fusion module is of a four-branch parallel three-dimensional convolution structure, and a BN layer and a ReLU activation function are connected in series behind each convolution layer and are used for extracting multi-scale space features; The characteristic splicing and fusing unit is used for splicing and fusing the preliminary time sequence characteristic and the multi-scale space characteristic in a channel dimension; The post ConvGRU is used for carrying out dynamic modeling on time sequence, and enhancing time sequence dependency characterization on the sea surface temperature evolution process.
3. The method for forecasting the sea surface temperature in two branches and two stages based on multi-element input according to claim 2, wherein the multi-scale space information fusion module comprises four branches: The first branch adopts a1 multiplied by 1 three-dimensional convolution layer for capturing small-scale features and retaining high-resolution details; the second branch adopts a3 multiplied by 3 three-dimensional convolution layer for enhancing the local feature characterization capability and expanding the channel dimension; The third branch is sequentially connected with a1 multiplied by 1 three-dimensional convolution layer and a 3 multiplied by 3 three-dimensional convolution layer in series, and the mesoscale features are firstly dimension-expanded and then extracted; The fourth branch is sequentially connected with a1 multiplied by 1 three-dimensional convolution layer and two cascaded 3 multiplied by 3 three-dimensional convolution layers in series to construct a large-scale receptive field to capture complex spatial modes.
4. The method for forecasting the sea surface temperature in two stages based on multi-element input according to claim 1, wherein the middle-long-term forecasting branch comprises an adaptive multi-scale space-time information fusion module, an encoder and a decoder which are connected in sequence; the self-adaptive multi-scale space-time information fusion module is used for carrying out channel self-adaptive weighting on the input multi-element space-time sequence, extracting multi-scale space features and time sequence dynamic features through the multi-scale space information fusion module and ConvGRU respectively, and carrying out fusion in channel dimensions to obtain space-time fusion features; The encoder is used for superposing the space-time fusion features, performing space-time encoding to obtain enhanced space-time fusion features, flattening the two-dimensional space features corresponding to the enhanced space-time fusion features in each time step into one-dimensional feature vectors through a remolding layer to obtain cross-time and cross-space feature sequences, unifying the channel dimensions of the feature sequences to preset embedding dimensions through a full-connection layer, inputting the feature sequences with unified dimensions to a transducer encoder, and outputting high-level space-time characterization after being processed by the transducer encoder with preset layers; and the decoder is used for extracting the characteristic vector of the final time step from the high-level space-time characterization, reconstructing the characteristic vector into a two-dimensional space structure and obtaining a single-channel sea surface temperature prediction result by regression.
5. The method for forecasting sea surface temperature in two branches and two stages based on multi-element input according to claim 4, wherein in the adaptive multi-scale spatial-temporal information fusion module, trainable channel weights are introduced into the input tensor X before entering the multi-scale spatial information fusion module and ConvGRU And weight distribution by applying non-negative constraints The contribution ratio of each ocean-atmosphere variable is dynamically adjusted, and the formula is as follows: ; wherein, reLU represents ReLU activation function, c is input variable number, namely channel number; The weight of channel c; the weighted feature X' is expressed as: ; Wherein, the Representing an input multi-element tensor; representing the weighted post tensor t moment in time Element values at h represents the number of weft grid points, and w represents the number of warp grid points; The output of the adaptive multi-scale spatio-temporal information fusion module is as follows: ; Wherein, the Representing the output of the adaptive multi-scale spatio-temporal information fusion module; representing a multi-scale spatial information fusion module, Representing a splice operation along the channel dimension.
6. The method for forecasting the sea surface temperature in two branches based on multi-element input according to claim 4, wherein the space-time coding is formed by overlapping time coding and space position coding, and is used for enhancing the distinguishing capability of different time steps and different space positions, and the formula is as follows: ; ; Wherein, the For initial space-time coding; Is time coded; Is spatially encoded; representing the output of the adaptive multi-scale spatio-temporal information fusion module, For final space-time coding.
7. The method for predicting sea surface temperature in two branches and two stages based on multi-element input according to claim 4, wherein the transducer encoder comprises a multi-head self-attention mechanism and a feedforward neural network layer; The multi-head self-attention mechanism is used for generating query vectors, key vectors and value vectors respectively through linear projection on an input feature sequence, dividing attention calculation into a plurality of parallel attention heads, calculating scaling dot product attention in each attention head, splicing the outputs of all the attention heads along feature dimensions, and mapping through an output projection matrix to obtain attention output; The feedforward neural network layer is used for sequentially carrying out dimension expansion, activation function nonlinear transformation and mapping of the attention output to the preset embedding dimension through the first full-connection layer to obtain feedforward output; the multi-head self-attention mechanism and the feedforward neural network layer are respectively combined with residual connection and layer normalization to stabilize the training process and retain original characteristic information.
8. The method for predicting sea surface temperature in two branches and two stages based on multi-element input according to claim 4, wherein the decoder is specifically configured to: receiving a high-level space-time representation of the encoder output; extracting a characteristic vector of a final time step in the high-level space-time characterization through a Lambda layer to serve as a state characteristic vector; The state feature vectors are sequentially input to a full-connection layer for dimension expansion, a remodelling layer is restored to a two-dimensional space structure, a1 multiplied by 1 convolution layer is regressed, and finally a sea surface temperature forecasting result of a single channel is output, wherein the decoder has the following formula: ; in the formula, Representing a final single-channel sea surface temperature forecasting result; Represents a remodelling layer; a feature vector representing a final time step obtained by Lambda layer processing; represented as a full connection layer weight matrix, Representing a bias term; Representing a1 x 1 convolution.
9. The method for forecasting the sea surface temperature in two branches and two stages based on multi-element input according to claim 1, wherein the first time period is 1-10 days, and the second time period is 11-30 days.
10. The method for forecasting the sea surface temperature in two stages based on multi-element input according to claim 1, wherein the steps S3 and S4 are characterized in that an autoregressive rolling forecasting mode is adopted to generate sea surface temperature forecasting results for a plurality of days in the future, and the method specifically comprises the following steps: Using time series data of a first preset day as input to forecast SST results of a future day; adding the SST result of the future day obtained by forecasting to the tail end of the time series data, and removing the data of the first day in the time series data to form new time series data of a first continuous preset day; taking the new time sequence data of the first preset day as input, and forecasting SST results of the next day; Repeating the steps, and repeating the loop iteration until SST forecasting results of a second preset day in the future are generated.

Description

Double-branch two-stage sea surface temperature forecasting method based on multi-element input Technical Field The invention relates to the technical field of marine environment forecasting, in particular to a double-branch two-stage marine surface temperature forecasting method based on multi-element input. Background The ocean is a resource pool and ecological base stone supporting sustainable development of human beings. Sea surface temperature (Sea Surface Temperature, SST) is a key parameter characterizing sea-air energy exchange, and minor anomalies thereof can significantly affect global climate, natural disasters and marine ecosystems. The accurate forecast SST not only can improve disaster prevention and reduction capability, but also can provide scientific decision basis for marine economic activities such as fishery, shipping and the like, and has great significance in reasonable development of marine resources, guarantee of marine rights and interests and promotion of high-quality development. Currently, SST forecasting methods can be mainly classified into numerical mode methods based on physical mechanisms and data driving methods based on historical observations. The numerical mode method is based on the dynamics and thermodynamic equation of the ocean-atmosphere system, and the energy and substance exchange process is simulated by solving a complex partial differential equation, so that the method has definite physical mechanism support. However, this type of method is computationally expensive, highly sensitive to initial conditions and parameter settings, and its forecast root mean square error remains between 0.6 ℃ and 1.0 ℃ for a long period of time, with significant bottlenecks facing precision improvement. In contrast, the data driving method constructs the forecast model by mining the statistical rules in the historical observation data, so that the calculation efficiency is high, and the realization path is flexible. However, when the traditional data driving algorithm (such as a Markov model, a Bayesian algorithm and the like) faces ocean data with high dimension, strong nonlinearity and strong space-time coupling, the feature extraction capability is limited, the generalization performance of the model is insufficient, and the requirement of high-precision prediction is difficult to meet. With the rapid development of deep learning technology, the method has obvious advantages in the aspects of automatic feature extraction, complex nonlinear modeling and time-space pattern recognition, and provides a new technical path for breaking through the SST forecasting precision bottleneck. SST intelligent forecasting methods based on convolutional neural networks, cyclic neural networks and variants thereof have gradually become important research hotspots in the field of intersection of ocean science and artificial intelligence. However, due to the complex structure of the monsoon system and ocean current and the interaction of multiple scales, the ocean environment is highly unstable, and the conventional deep learning method still faces the problem of insufficient precision and adaptability. The prior research focuses on SST forecast of a single variable, ignores the coupling relation between SST and offshore air temperature (T2 m), sea level air pressure (MSL) and other ocean-atmosphere elements, and therefore limits the depicting capability of the model on complex physical processes. In addition, when the existing method is used for coping with medium-long-term forecasting tasks, the problem of error accumulation is common, so that the forecasting precision is obviously reduced along with the increase of forecasting time, and the requirements of actual business application on medium-long-term forecasting stability and reliability are difficult to meet. Therefore, it is necessary to further explore a deep learning-based ocean-atmosphere multi-element joint prediction method, introduce a attention mechanism and a high-efficiency feature extraction structure by fusing multi-source multi-variable information, construct an SST prediction framework considering global evolution rules and regional local features, and design differentiated modeling strategies aiming at different characteristics of short-term prediction and medium-term prediction so as to effectively inhibit error accumulation and realize cooperative improvement of SST prediction precision and stability. Disclosure of Invention In order to solve the technical problems, the invention provides a double-branch two-stage sea surface temperature forecasting method based on multi-element input, so as to achieve the purposes of cooperatively improving sea surface temperature forecasting precision and medium-long-term forecasting stability and effectively inhibiting error accumulation in autoregressive rolling forecasting. In order to achieve the above purpose, the technical scheme of the invention is as follows: a double-branch two-st