Water level prediction method and device based on multi-modal fusion and double-flow space-time network PredFormer

Abstract

The invention discloses a water level prediction method and device based on multi-modal fusion and a double-flow space-time network PredFormer, and the method comprises the steps: firstly, building a double-flow space-time network PredFormer model frame in a dynamic manner from a time sequence signal prediction problem as an entry point and considering that the water level and meteorological data have high nonlinearity and dynamic space-time dependence, and carrying out the prediction of a time sequence signal; obtaining spatio-temporal characteristics of rainfall data and water level data of a plurality of meteorological stations; thirdly, performing feature-level fusion on the obtained water level data time sequence features and multi-site rainfall data spatial-temporal features by using an efficient low-rank tensor fusion method; and finally, a multi-layer perceptron is used as an output prediction result. According to the method, the nonlinear time sequence relation problem of water level prediction is considered, time-space dynamic association of water level data and rainfall data can be fused, and meanwhile, the cross-site prediction problem of personality and generality of the water level data of the hydrometric station and the rainfall data of different meteorological stations is considered. Therefore, the method has strong generalization ability in water level prediction across multiple meteorological stations.

Description

Technical Field

[0001] The present invention relates to the field of water level prediction, and more specifically, to a water level prediction method and device based on a multi-modal and spatio-temporal full attention mechanism PredFormer. Background Art

[0002] The prediction of water level and water regime can timely predict information such as river water level and flood peak flow, which is very important for flood control and disaster reduction, production and life, power generation scheduling, etc. Based on the prediction, the industrial and agricultural production plans and domestic water intake plans can be reasonably arranged, and the power generation plans of hydropower stations can also be reasonably arranged to make full use of water resources and improve the power generation efficiency of hydropower stations. The prediction of water level can timely detect natural disasters such as possible floods and droughts, so as to take corresponding disaster prevention and relief measures to reduce losses.

[0003] There are many existing data-driven water level prediction methods. However, due to the highly non-linear and dynamic spatio-temporal dependence of water level data, timely and accurate prediction of water level data is still an important challenge faced by the academic community and management departments. In addition, although there are some methods using neural network models for water level prediction, these methods do not consider the relationship between water level stations and multiple meteorological stations, as well as the data relationships between different meteorological stations and within stations, resulting in low generalization ability of the model and the need to improve prediction accuracy. Summary of the Invention

[0004] Aiming at the technical problems existing in the prior art, the present invention provides a water level prediction method based on multi-modal fusion and a two-stream spatio-temporal network PredFormer. First, starting from the problem of time series signal prediction, considering the highly non-linear and dynamic spatio-temporal dependence between water level and meteorological data, a two-stream spatio-temporal network Predformer model framework is built in a dynamic manner to obtain the spatio-temporal characteristics of precipitation data and water level data of multiple meteorological stations; secondly, the Efficient Low-rank Multimodal Fusion (ELMF) method is used to perform feature-level fusion on the time series characteristics of the water level data obtained above and the spatio-temporal characteristics of multi-station precipitation data; finally, a Multilayer Perceptron (MLP) is used to output the prediction result. This method not only considers the non-linear time series relationship problem of water level prediction, but also can fuse the spatio-temporal dynamic association of water level data and precipitation data, and at the same time takes into account the cross-station prediction problems of the individuality and commonality of water level data of hydrological stations and precipitation data of different meteorological stations. This makes the method have strong generalization ability in the water level prediction across multiple meteorological stations.

[0005] To achieve the above object, the first aspect of the present invention provides a water level prediction method based on multi-modal fusion and a two-stream spatio-temporal network PredFormer, including:

[0006] Determine the corresponding meteorological station according to the hydrological station to be predicted;

[0007] Collect the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station;

[0008] Preprocess the collected water level monitoring data and precipitation monitoring data;

[0009] Build a water level prediction model based on the dual-stream spatio-temporal network PredFormer, and use the preprocessed data to train the water level prediction model based on the dual-stream spatio-temporal network PredFormer. Among them, the water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch inputs the embedding, positional encoding and PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism, and successively establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations to obtain the spatio-temporal features of precipitation data of multiple meteorological stations. The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embedding, positional encoding and PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism to obtain the temporal features of the water level data; The feature fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the obtained spatio-temporal features of precipitation data of multiple meteorological stations and the temporal features of water level data to obtain the fused features; The multi-layer perceptron outputs the water level prediction result based on the fused features;

[0010] Use the water level prediction model based on the dual-stream spatio-temporal network PredFormer to predict the water level of the hydrological station to be predicted.

[0011] In one implementation, determining the corresponding meteorological station according to the hydrological station to be predicted includes:

[0012] Based on the DEM, divide the river basin range where the hydrological station to be predicted is located, and select the meteorological stations located upstream of the hydrological station and within the river basin range as the corresponding meteorological stations.

[0013] In one implementation, preprocessing the collected water level monitoring data and precipitation monitoring data includes:

[0014] Perform missing value, outlier and standardization processing on the collected water level monitoring data and precipitation monitoring data.

[0015] In one implementation, before the PredFormer model with spatio-temporal full attention mechanism in the first branch inputs embeddings, positional encodings, and the PredFormer encoder to successively establish the spatial relationships between weather stations and the temporal relationships within weather stations, and obtain the spatio-temporal features of precipitation data of multiple weather stations, the method further includes:

[0016] Use a graph convolutional neural network to model the spatial position relationships of multi-station precipitation data, construct the association information between stations, and obtain multi-station structured graph data.

[0017] In one implementation, the PredFormer model with spatio-temporal full attention mechanism in the first branch inputs embeddings, positional encodings, and the PredFormer encoder to successively establish the spatial relationships between weather stations and the temporal relationships within weather stations, and obtain the spatio-temporal features of precipitation data of multiple weather stations, including:

[0018] Use the input embeddings of the PredFormer model with spatio-temporal full attention mechanism to take the input multi-station structured graph data as multiple input sub-channels. The data of each sub-channel is flattened into a one-dimensional vector, and then the one-dimensional vector is projected into the hidden dimension through a linear layer to obtain the corresponding tensor;

[0019] Use absolute positional encoding for the positional encoding of the spatial positions and time step sequences of each station;

[0020] Through multiple stacked gated transformer blocks of the PredFormer encoder combined with the multi-head self-attention mechanism, obtain the spatio-temporal features of precipitation data of multiple weather stations based on the positional encoding.

[0021] In one implementation, the second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embeddings, positional encodings, and the PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism, and obtains the temporal features of the water level data, including:

[0022] Use the input embeddings of the PredFormer model with spatio-temporal full attention mechanism to process the input single-station water level data to obtain the corresponding tensor;

[0023] Use absolute positional encoding for the positional encoding of the single-station water level data;

[0024] Through multiple stacked gated transformer blocks of the PredFormer encoder combined with the multi-head self-attention mechanism, obtain the temporal features of the water level data based on the positional encoding.

[0025] In one embodiment, the output layer of the efficient low-rank tensor fusion method with modality-specific factors employs an MLP regression layer.

[0026] Based on the same inventive concept, the second aspect of the present invention provides a water level prediction device based on a multi-modal and spatio-temporal full attention mechanism PredFormer, comprising:

[0027] A meteorological station determination module, configured to determine the corresponding meteorological station according to the hydrological station to be predicted;

[0028] A data collection module, configured to collect the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station;

[0029] A data preprocessing module, configured to preprocess the collected water level monitoring data and precipitation monitoring data;

[0030] A model construction and training module, configured to construct a water level prediction model based on the dual-stream spatio-temporal network PredFormer, and train the water level prediction model based on the dual-stream spatio-temporal network PredFormer by using the preprocessed data. The water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module, and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch inputs through the PredFormer model of the spatio-temporal full attention mechanism, position encoding, and the PredFormer encoder, successively establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations, and obtains the spatio-temporal features of precipitation data of multiple meteorological stations. The second branch processes the temporal sequence relationship of the water level data of the hydrological station to be predicted through the PredFormer model of the spatio-temporal full attention mechanism, position encoding, and the PredFormer encoder, and obtains the temporal sequence features of the water level data. The feature fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the obtained spatio-temporal features of precipitation data of multiple meteorological stations and the temporal sequence features of the water level data, and obtains the fused features. The multi-layer perceptron outputs the water level prediction result based on the fused features;

[0031] A water level prediction module, configured to use the trained water level prediction model based on the dual-stream spatio-temporal network PredFormer for the hydrological station to be predicted

[0032] Based on the same inventive concept, the third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the water level prediction method based on multi-modal fusion and the dual-stream spatio-temporal network PredFormer described in the first aspect.

[0033] Based on the same inventive concept, a fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the water level prediction method based on multimodal fusion and the dual-stream spatio-temporal network PredFormer described in the first aspect.

[0034] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0035] The present invention provides a water level prediction method based on multimodal fusion and the dual-stream spatio-temporal network PredFormer, and constructs a water level prediction model based on the spatio-temporal network Predformer. Starting from the time series signal prediction problem, considering that the water level and meteorological data (precipitation monitoring data) have highly non-linear and dynamic spatio-temporal dependencies, a Predformer model framework is constructed in a dynamic manner. Through the spatio-temporal full attention mechanism PredFormer model, the spatio-temporal features of precipitation data from multiple meteorological stations and the time series features of water level data are dynamically obtained; and the efficient low-rank tensor fusion (ELMF) method is used to perform feature-level fusion of the time series features of the extracted water level data and the spatio-temporal features of multi-site precipitation data; finally, a multilayer perceptron (MLP) is used to output the prediction result. This method not only considers the non-linear time series relationship problem of water level prediction, but also can fuse the spatio-temporal dynamic association of water level data and precipitation data, and at the same time takes into account the cross-site prediction problem of the individuality and commonality of water level data of hydrological stations and precipitation data of different meteorological stations, so that the prediction model has strong generalization ability in cross-multiple meteorological station water level prediction and improves the accuracy of water level prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0037] Figure 1 It is a flowchart of the water level prediction method based on multimodal fusion and the dual-stream spatio-temporal network PredFormer in the embodiments of the present invention.

[0038] Figure 2 It is a schematic diagram of a meteorological station in the embodiments of the present invention.

[0039] Figure 3It is a comparison chart of the prediction results between the method of the embodiment of the present invention and the existing method.

[0040] Figure 4 It is a module diagram of the water level prediction device based on multimodal fusion and the dual-stream spatio-temporal network PredFormer in the embodiment of the present invention.

[0041] Figure 5 It is a structural diagram of the computer-readable storage medium provided by the embodiment of the present invention.

[0042] Figure 6 It is a structural diagram of the computer device in the embodiment of the present invention. Specific embodiments

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] Embodiment 1

[0045] This embodiment discloses a water level prediction method based on multimodal fusion and the dual-stream spatio-temporal network PredFormer. Please refer to Figure 1 , including:

[0046] S1: Determine the corresponding meteorological station according to the hydrological station to be predicted;

[0047] S2: Collect the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station;

[0048] S3: Preprocess the collected water level monitoring data and precipitation monitoring data;

[0049] S4: Build a water level prediction model based on the dual-stream spatio-temporal network PredFormer, and use the preprocessed data to train the water level prediction model based on the dual-stream spatio-temporal network PredFormer. The water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module, and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch inputs embeddings, positional encodings, and the PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism, and successively establishes the spatial relationships between meteorological stations and the temporal relationships within meteorological stations to obtain the spatio-temporal features of precipitation data of multiple meteorological stations. The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embeddings, positional encodings, and the PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism to obtain the temporal features of the water level data. The feature fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the obtained spatio-temporal features of precipitation data of multiple meteorological stations and the temporal features of water level data to obtain the fused features. The multi-layer perceptron outputs the water level prediction result based on the fused features;

[0050] S5: Use the trained water level prediction model based on the dual-stream spatio-temporal network PredFormer to predict the water level of the hydrological station to be predicted.

[0051] Specifically, S1 is the basin division and the selection of meteorological stations, S2 is the collection of precipitation data of multiple meteorological stations and the water level data of hydrological stations, S3 is the data preprocessing, and S4 is the model construction. In the present invention, a water level prediction model based on the dual-stream spatio-temporal network PredFormer is constructed. The feature extraction module adopts a two-branch structure, and each branch uses the PredFormer model with spatio-temporal full attention mechanism to obtain the spatio-temporal features of precipitation data of multiple meteorological stations and the temporal features of water level data respectively. The feature fusion module adopts an efficient low-rank tensor fusion method with modality-specific factors, so as to strengthen the correlation of multi-modal features during the model learning process and improve the prediction accuracy compared with other fusion methods. S5 is to use the constructed model to predict the water level. Since the water level prediction model not only considers the non-linear temporal relationship problem of water level prediction, but also can fuse the spatio-temporal dynamic association of water level data and precipitation data, and at the same time takes into account the cross-station prediction problem of the individuality and commonality of hydrological station water level data and precipitation data of different meteorological stations. The prediction model has strong generalization ability in the water level prediction across multiple meteorological stations, so accurate prediction results can be obtained.

[0052] In one implementation, determining the corresponding meteorological stations according to the hydrological station to be predicted includes:

[0053] Based on the DEM, divide the river basin range where the hydrological station to be predicted is located, and select the meteorological stations located upstream of the hydrological station and within the river basin range as the corresponding meteorological stations.

[0054] In specific implementation, according to the hydrological station for which the water level is to be predicted, divide the river basin range where the hydrological station is located based on the DEM, and then select the meteorological stations located upstream of the hydrological station and within the river basin range as the backup meteorological stations. There is no limit on the number of meteorological stations, generally not less than 3.

[0055] In one implementation manner, preprocess the collected water level monitoring data and precipitation monitoring data, including:

[0056] Perform missing value, outlier, and standardization processing on the collected water level monitoring data and precipitation monitoring data.

[0057] In specific implementation, the preprocessing method is as follows:

[0058] First, adopt methods such as mean replacement method, linear interpolation, time interpolation, filling interpolation, nearest neighbor interpolation, etc., and select the best missing value processing method according to the fitting effect.

[0059] Second, use the 3σ criterion and Grubbs criterion to process outliers. Assume that the measured variable is measured with equal precision, and x1, x2,..., x i , are obtained, calculate its arithmetic mean and the residual error and calculate the standard error σ according to the Bessel formula. If the residual error v b of a certain measured value x b (1 ≤ b ≤ n) satisfies then x b is considered a bad value containing a gross error value and is marked. The Bessel formula is as follows:

[0060]

[0061] n represents the number of measured variables.

[0062] Third, use linear normalization ("max-min" normalization) to obtain the difference between the maximum and minimum values in the two types of data sets, and normalize the data points to a base number.

[0063]

[0064] x′ is the normalized data, x is the original data, max(x) and min(x) are the maximum and minimum values of the data respectively.

[0065] In one implementation, before the PredFormer model with spatio-temporal full attention mechanism in the first branch inputs the embedding, positional encoding, and PredFormer encoder, and successively establishes the spatial relationships between weather stations and the temporal relationships within weather stations to obtain the spatio-temporal features of precipitation data of multiple weather stations, the method further includes:

[0066] Model the spatial position relationships of multi-station precipitation data using a graph convolutional neural network, construct the association information between stations, and obtain multi-station structured graph data.

[0067] Specifically, since the multi-station precipitation data shows a sparse distribution and cannot be directly input into the PredFormer for processing. Therefore, in this implementation, a convolutional neural network is used to perform spatio-temporal encoding on the precipitation data, convert it into structured graph data, and then input it into the PredFormer model. The spatial dependence is captured through the GCN to solve the data sparsity problem.

[0068] In one implementation, the PredFormer model with spatio-temporal full attention mechanism in the first branch inputs the embedding, positional encoding, and PredFormer encoder, and successively establishes the spatial relationships between weather stations and the temporal relationships within weather stations to obtain the spatio-temporal features of precipitation data of multiple weather stations, including:

[0069] Use the embedding of the PredFormer model with spatio-temporal full attention mechanism to take the input multi-station structured graph data as multiple input sub-channels. The data of each sub-channel is flattened into a one-dimensional vector, and then the one-dimensional vector is projected into the hidden dimension through a linear layer to obtain the corresponding tensor;

[0070] Use absolute positional encoding for positional encoding to encode the spatial positions and time step sequences of each station;

[0071] Through multiple stacked gated transformer blocks of the PredFormer encoder combined with the multi-head self-attention mechanism, obtain the spatio-temporal features of precipitation data of multiple weather stations based on the positional encoding.

[0072] Specifically, the dual-stream PredFormer spatio-temporal prediction model uses two PredFormer models to extract spatio-temporal features of multi-station precipitation data and target station water level data respectively. The specific principle of the PredFormer model is as follows:

[0073] One is the PredFormer model input embedding: for the precipitation data of multiple stations, the data of multiple stations are used as multiple input sub-channels. The data of each sub-channel is flattened into a one-dimensional (1D) vector, and then these one-dimensional vectors are projected into the hidden dimension D through a linear layer to obtain a tensor Where B, T, N, and D represent the number of batches, time steps, number of stations, and dimensions respectively; for the water level data of a single station, it is first flattened into a one-dimensional (1D) vector and then projected into the hidden dimension D through a linear layer to obtain a tensor

[0074] Second, PredFormer positional encoding: Absolute positional encoding is adopted, and the spatial positions of each station and the time step sequence are encoded through sine functions respectively.

[0075] Specifically, for position k, its positional encoding e k The i-th component is calculated as follows:

[0076]

[0077] Each dimension component of the triangular positional encoding is actually a sine function (even dimension components) or cosine function (odd dimension components) of position k, and sine and cosine functions are used respectively according to the parity of the dimension.

[0078] Third, the PredFormer encoder: The PredFormer encoder is stacked by gated transformer blocks in various ways.

[0079] Gated linear units (GLUs) are usually used to replace simple linear transformations, involving the element-wise product of two linear projections, where one projection passes through the Sigmoid function. Various GLU variants control the information flow by replacing Sigmoid with other non-linear functions. For example, SwiGLU replaces Sigmoid with the Swish activation function (SiLU), and the calculation formula is as follows:

[0080] Swish β (x) = xσ(βx)

[0081]

[0082] Where x is the hidden representation at a specific position in the sequence, β is an additional parameter, usually taken as 1; W and V are weight matrices, and b and c are bias vectors, represents the outer product of tensors.

[0083] The gated transformer block (GTB) combines the multi-head self-attention mechanism (MSA) and the feed-forward neural network (FFN) based on SwiGLU. The GTB is defined as:

[0084] Y l= MSA(LN(Z l )) + Z l

[0085] Z l+1 = SwiGLU(LN(Y l )) + Y l

[0086] where MSA represents multi - head self - attention, LN is layer normalization, Z l , Y l both represent the features of the l - th layer.

[0087] Each element of the MSA input sequence is first mapped to three vectors, which is usually achieved by a linear transformation with three weight matrices. Specifically, the input sequence X is multiplied by the respective weight matrices to obtain Q, K, and V. MSA is defined as:

[0088] MultiHead(Q, K, V) = Concat(head1, …, head h )W O

[0089] head i = Attention(QW i Q , KW i K , VW i V )

[0090] where MultiHead(Q, K, V) represents the output sequence, head i represents the multi - head attention weight of the i - th element, Q, K, V respectively represent the three vectors that each element of the input sequence is mapped to, namely the query (the target to be analyzed), the key (the element to be analyzed), and the value (the result to be analyzed) matrices, W i Q , W i K , W i V are the respective corresponding weights.

[0091] To convert the attention scores to weights, the Softmax function is applied for normalization, and an N×N attention weight matrix is calculated, indicating the influence degree of each patch on other patches. Softmax ensures that the sum of all output weights is 1, so that the model can learn the importance of each element pair:

[0092]

[0093] Among them, Attention(Q, K, V) represents the multi-head attention weights of the output elements, and QK T is the dot product between the query and the key, indicating the degree of attention of each query to all keys, is the scaling factor used to stabilize gradient propagation.

[0094] The first stream (the first branch) first uses a graph convolutional neural network (GCN) to model the spatial position relationship of multi-site precipitation data, then extracts the cross-site spatial relationship and the intra-site temporal dependence relationship through the PredFormer spatio-temporal prediction model, and finally maps it to the spatio-temporal features of multi-site precipitation data through a multi-layer perceptron (MLP). The specific principle is as follows:

[0095] First, use GCN to model the spatial position relationship of multi-site precipitation data and construct the association information between sites. Let the precipitation data be represented as a graph signal on the site set GCN extracts spatial information through the graph Laplacian operator The basic calculation method is:

[0096]

[0097] where U1 (0) = X1, that is, the input multi-site precipitation data; W (l) is the training weight matrix of the L-th layer; is the normalized Laplacian matrix, is the adjacency matrix with self-loops, is the degree matrix; σ(·) is the ReLU activation function.

[0098] Subsequently, extract the cross-site spatial relationship and the intra-site temporal dependence relationship through the PredFormer spatio-temporal prediction model. Let the spatial features output by GCN be Take it as the input of the PredFormer model, and model the spatio-temporal relationship of the input U1 through the PredFormer model to output the prediction result:

[0099]

[0100] where is the output feature of PredFormer. Finally, use MLP to perform a non-linear transformation on the output of PredFormer and map it to the spatio-temporal features of multi-site precipitation data:

[0101]

[0102] where is to The two-dimensional vector after flattening the dimensions N and D, σ(·) is the ReLU activation function, ReLU(X) = max(0, X); W1 and W2 are the trained weight matrices, and b1 and b2 are the trained bias vectors; are the spatio-temporal characteristics of multi-site precipitation data.

[0103] In one implementation, the second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embedding, positional encoding, and PredFormer encoder of the PredFormer model with spatio-temporal full attention mechanism, and obtains the temporal characteristics of the water level data, including:

[0104] Process the input single-site water level data through the input embedding of the PredFormer model with spatio-temporal full attention mechanism to obtain the corresponding tensor;

[0105] Perform positional encoding on the single-site water level data by using absolute positional encoding;

[0106] Through multiple stacked gated transformer blocks of the PredFormer encoder combined with the multi-head self-attention mechanism, obtain the temporal characteristics of the water level data based on the positional encoding.

[0107] In the specific implementation process, the second stream (second branch) uses the PredFormer spatio-temporal prediction model to extract the temporal characteristics of the water level data of the target hydrological station, and further maps it through the MLP to obtain the temporal characteristics of the water level data. The specific principle is as follows:

[0108] Let the water level data of the target hydrological station be a time series The PredFormer model learns the long-term dependence relationship of the time series through the self-attention mechanism, and its input-output relationship can be expressed as:

[0109]

[0110] where is the PredFormer output feature. Further use the MLP to perform a non-linear transformation on the output of the PredFormer and map it to the temporal characteristics of the water level data:

[0111]

[0112] where is the temporal characteristic of the water level data.

[0113] In one implementation, the output layer of the efficient low-rank tensor fusion method with modality-specific factors uses an MLP regression layer.

[0114] The efficient low-rank multi-modal fusion (ELMF) method is used to fuse the multi-site precipitation spatio-temporal feature V1 and the water level data time series feature V2 obtained in step S4 to obtain a fusion feature, and its input-output relationship can be expressed as:

[0115] H = ELMF(V1, V2)

[0116] where is the fusion feature.

[0117] The specific principle of the ELMF method is as follows:

[0118] The multi-modal fusion is expressed as a multi-linear function f:

[0119] V1, V2,..., VM → H

[0120] where is the vector space of the input modality, and H is the output vector space.

[0121] Given a set of vector representations, encodes the unimodal information of M different modalities. The goal of multi-modal fusion is to integrate the unimodal representations into a compact multi-modal representation for downstream tasks.

[0122] The input tensor formed by the unimodal representations is calculated by the following formula:

[0123]

[0124] where, represents the tensor outer product on a set of vectors indexed by m, and zm is the input representation appended with 1s.

[0125] The input tensor then generates a vector representation through the linear layer g(·):

[0126]

[0127] where W is the weight of this layer and b is the bias. Since is a tensor of order M (where M is the number of input modalities), the weight W is naturally a tensor of order (M + 1) in; the additional (M + 1) dimensions correspond to the size of the output representation dh. In the tensor dot product the weight tensor W can be regarded as a tensor of order M of dh.

[0128] For a tensor of order M there always exists an exact decomposition into the form of vectors:

[0129]

[0130] Among them are called the rank-R decomposition factors of the original tensor.

[0131] Recombine and concatenate these vectors into M modality-specific low-rank factors. Let Then for modality m, is its corresponding low-rank factor, and the low-order weight tensor can be recovered by the following method:

[0132]

[0133] Therefore, h can be calculated by the following formula:

[0134]

[0135] Note that for all the sizes of the second dimensions are the same. We define their outer product to cover only the non-shared dimensions:

[0136] By introducing the low-rank factors, it is necessary to calculate for further calculation. Using the tensor naturally decomposes into the original input which is parallel to the modality-specific low-rank factors.

[0137] Using simplifying the equation gives:

[0138]

[0139] Among them represents the element-wise product on the tensor sequence:

[0140] Using the water level prediction model based on the two-stream spatio-temporal network PredFormer, the water level of the hydrological station to be predicted is predicted. Using the fused feature H obtained in step five of the MLP decoding, the future water level prediction result is output. Similar to the steps in S4 above, the spatio-temporal features of precipitation data from multiple meteorological stations and the temporal features of water level data are respectively extracted through the feature extraction module (two-stream spatio-temporal network PredFormer); and the extracted features are fused at the feature level using the efficient low-rank tensor fusion method; finally, a multi-layer perceptron (MLP) is used as the output to predict the result.

[0141] The specific processing process of the multi-layer perceptron is as follows:

[0142] H f = Flatten(H)

[0143]

[0144] wherein is the one-dimensional vector after flattening H, is the predicted value of the future water level.

[0145] In one embodiment, the method further includes a post-evaluation of the accuracy of the water level prediction model. According to the water level prediction results of multimodal fusion and the dual-stream spatio-temporal network PredFormer, and subsequent actual water level monitoring data, the root mean square logarithmic error (RMSLE) and weighted mean absolute percentage error (WMAPE) are used as evaluation indicators to perform a post-evaluation of the accuracy of the water level prediction model.

[0146]

[0147] where y i and respectively represent the actual value and the predicted value of the water level.

[0148] This index increases or decreases the penalty for small predicted values under the same root mean square error, thereby reducing the risk of flood control impact caused by small short-term water level predictions.

[0149]

[0150] where y i and respectively represent the actual value and the predicted value of the water level.

[0151] WMAPE weights the value impact of a single data pair on the relevant population, and the error fluctuations caused by extreme values are small, and its applicability is better than the mean absolute percentage error (MAPE).

[0152] Next, the concept and key points of the present invention will be described.

[0153] In terms of water level changes, the water level itself has a certain continuity in time. How to capture this temporal relationship is the basis of water level prediction. Secondly, the water level value itself is correlated with meteorological factors (especially precipitation). How to effectively and efficiently capture the spatial and temporal dependencies between water level data and meteorological data is very crucial. There are many existing data-driven water level prediction methods. However, due to the highly non-linear and dynamic spatio-temporal dependence of water level data, accurately predicting water level data in a timely manner is still an important challenge faced by the academic community and management departments. Compared with general spatio-temporal prediction tasks, the particularity of water level prediction lies in multi-variable input (water level and multiple precipitation data) and the difference in the spatial characteristics of data (precipitation has spatio-temporal characteristics, while water level only has time characteristics). Therefore, the existing spatio-temporal network PredFormer model cannot be directly used for water level prediction.

[0154] Based on the demand for water level prediction, the following improvements have been made to the spatio-temporal network PredFormer model in the present invention:

[0155] 1. Adapt to multi-variable input: Adopt a dual-branch structure, use PredFormer to encode water level features and precipitation features respectively, and then fuse the two features for prediction;

[0156] 2. Handle the sparsity of precipitation data: Use a graph convolutional neural network (GCN) to encode multi-site precipitation data, capture its spatial dependence, and solve the data sparsity problem;

[0157] 3. Optimize feature extraction and compression: Replace the Patch Recovery of PredFormer with a fully connected layer to reduce the feature dimension, reduce the computational complexity, and make the model learning more efficient.

[0158] Meanwhile, in the prior art, the processing of multi-site precipitation data faces the following technical difficulties: 1. Data sparsity: The multi-site precipitation data shows a sparse distribution and cannot be directly input into PredFormer for processing. Therefore, it is necessary to perform spatio-temporal encoding on the precipitation data and convert it into structured graph data before it can be input into the prediction model; 2. Spatial flow relationship: The precipitation data has strong spatial mobility between different sites, that is, the change of precipitation has spatial correlation. Therefore, the geographical topological relationship between sites needs to be fully considered during the encoding process to effectively capture the spatial dependence between sites.

[0159] Regarding the fusion of features, the efficient low-rank tensor fusion (ELMF) method is usually used for classification tasks such as multi-modal sentiment recognition and visual question answering. Different from the classification tasks in the prior art, in the present invention, when applying it to the regression task of water level prediction, the following improvements need to be made:

[0160] 1. Output layer modification: Replace the ELMF output layer with an MLP regression layer to adapt to the continuous prediction of water level;

[0161] 2. Feature encoding adjustment: When encoding multi-modal features, the spatio-temporal mobility of the data needs to be considered. Therefore, PredFormer is used for spatio-temporal feature encoding to capture the spatio-temporal dependence of precipitation and water level data. At the same time, considering the relationship between water level stations and multiple meteorological stations, as well as the spatio-temporal relationships between different meteorological stations and within stations, a water level prediction method based on multi-modal fusion and a dual-stream spatio-temporal network PredFormer is relatively rare.

[0162] In the prior art, CN116151411A proposed a water level prediction method based on the spatio-temporal fusion of multiple features of AutoFormer, and CN116562416B disclosed a reservoir water level prediction method based on the variational mode and delayed feature Transformer model. Compared with the AutoFormer and Transformer models in the above methods, the PredFormer model of the present invention can process time and space information simultaneously in one framework, does not require a computational complexity that is quadratic in the sequence length, has neither autoregression nor any convolution, is both simple and efficient, has fewer parameters, and its performance is significantly better than previous methods; in addition, the model architecture of this method uses two independent branches for feature extraction, adopts the spatio-temporal characteristics of multi-site precipitation, and adopts the multi-modal efficient low-rank tensor fusion ELMF. Using this fusion method can strengthen the correlation of multi-modal features during the model learning process and improve the prediction accuracy compared with other fusion methods.

[0163] The method proposed by the present invention is verified and described below through specific examples.

[0164] Example: Taking a hydrological station in the lower reaches of the Han River as an example, a water level prediction method based on multi-modal fusion and spatio-temporal full attention mechanism PredFormer is used to predict the water level data of the hydrological station within the next month.

[0165] Step 1: According to the site where the water level is to be predicted, divide the river basin range where the hydrological station is located based on DEM, extract the basin range controlled by the hydrological station, and select three basic meteorological stations A, B, and C located within this basin range.

[0166] Step 2: Collect the water level monitoring data (unit: m 3 / s) of this hydrological station from 2016 to 2020, and the monthly precipitation monitoring data (unit: mm) of meteorological stations A, B, and C during the same period.

[0167] Step 3: Preprocess the monitoring data of the hydrological monitoring station and the three meteorological stations, mainly including missing value processing, outlier processing, standardization, etc. The descriptive statistical results of the processed precipitation and water level data are as follows:

[0168] Table 1 Descriptive statistics of data

[0169] Statistical value A B C Hydrological station Sample size 444 444 444 444 Mean value 61.89 60.52 72.27 21.58 Standard deviation 54.22 53.54 68.22 2.59 Minimum value 0 0 0 17.67 25% quantile 15.94 20.26 24.13 19.55 Median 45.34 45.47 52.58 20.83 75% quantile 98.11 89.54 100.52 23.27 Maximum value 338.58 435.1 447.55 30.75

[0170] Step 4: Through the spatial position encoding of the GCN model, the input, position encoding, and encoder of the PredFormer model are used to establish the spatial relationship between stations and the time relationship within stations in sequence, and obtain the spatio-temporal characteristics of the precipitation data of the three meteorological stations. The specific model parameter settings are as follows:

[0171] Table 2 Model parameters and specific values for multi-site precipitation data

[0172] Model component Parameter Value Input data dimension (B, T, N, D) (1,12,3,1) Adjacency matrix A 3×3 normalized adjacency matrix Number of GCN layers L(GCN) 2 GCN hidden layer dimension D(GCN) 64 GCN activation function / ReLU PredFormer input dimension (B, T, N, D) (1,12,3,64) PredFormer positional encoding Method Absolute positional encoding Number of PredFormer layers L(PredFormer) 4 Number of attention heads h(MSA) 8 PredFormer hidden layer dimension D(PredFormer) 64 PredFormer activation function / SwiGLU MLP dimension D(MLP) 64 MLP activation function / ReLU Output data dimension (B, T, D) (1,12,64)

[0173] Step 5: Use the PredFormer model input, positional encoding, and encoder to process the temporal relationship of the water level data at the target hydrological station to obtain the temporal features of the water level data.

[0174] Table 3 Model parameters and specific values for the water level data of the target site

[0175]

[0176]

[0177] Step 6: Use the efficient low-rank tensor fusion (ELMF) method with modality-specific factors to perform feature-level fusion of the temporal features of the water level data obtained previously and the spatio-temporal features of the multi-site precipitation data, improve the correlation between features, and enhance the expressive ability of the model to obtain fused features. The fused features are used to output the water level prediction results for 2020 through an MLP.

[0178] Table 4 ELMF parameter information and specific values

[0179] Model component Parameter Value Input data dimension (B, T, N, D) (1,12,2,64) Low-rank decomposition dimension r 32 Output data dimension (B, T, D) (1,12,128) MLP dimension D(MLP) 64 MLP activation function / ReLU Output data dimension (B, T, D) (1,12,1)

[0180] Step 7: Select the data samples from 2016 to 2019 as the training set and the data in 2020 as the test set to verify the effect of the model. Based on the multi-layer perceptron (MLP), output the water level prediction results for 2020; at the same time, use traditional methods such as MLP, long short-term memory network (LSTM), gated recurrent unit (GRU), and spatio-temporal prediction network PredFormer to predict the prediction results of the hydrological station in 2020.

[0181] Step 8: According to the model prediction results, use the root mean square logarithmic error (RMSLE) and weighted mean absolute percentage error (WMAPE) as evaluation indicators to evaluate the performance of the water level prediction methods of multi-modal fusion and PredFormer spatio-temporal network compared with traditional prediction models. After calculation, the RMSLE and WMAPE of the prediction method of the present invention are 6.34% and 21.02% respectively, and the model prediction effect is better than that of traditional prediction methods.

[0182] Table 5 Comparison of water level prediction result indicators of different prediction models

[0183] Model RMSELE(%) WMAPE(%) MLP 9.94 35.01 LSTM 9.52 33.71 GRU 8.41 29.21 PredFormer 7.81 27.04 Multi-modal fusion + PredFormer 6.34 21.02

[0184] The water level prediction method based on multimodal fusion and the dual-stream spatio-temporal network PredFormer in this embodiment obtains the data sample data of hydrological stations and meteorological stations. On the basis of the preprocessing of these two types of time series data, it constructs the prediction model framework of the dual-stream spatio-temporal network PredFormer, and then based on the data fusion of efficient low-rank multimodal fusion, outputs the water level prediction result based on the multi-layer perceptron (MLP). The model has good prediction accuracy and strong generalization performance.

[0185] Embodiment 2

[0186] Based on the same inventive concept, this embodiment discloses a water level prediction device based on multimodal and spatio-temporal full attention mechanism PredFormer. Please refer to Figure 2 , including:

[0187] The meteorological station determination module 201 is used to determine the corresponding meteorological station according to the hydrological station to be predicted;

[0188] The data collection module 202 is used to collect the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station;

[0189] The data preprocessing module 203 is used to preprocess the collected water level monitoring data and precipitation monitoring data;

[0190] The model construction and training module 204 is used to construct a water level prediction model based on the dual-stream spatio-temporal network PredFormer, and use the preprocessed data to train the water level prediction model based on the dual-stream spatio-temporal network PredFormer. Among them, the water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch sequentially establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations through the input of the PredFormer model with spatio-temporal full attention mechanism, positional encoding and the PredFormer encoder, and obtains the spatio-temporal features of precipitation data of multiple meteorological stations. The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input of the PredFormer model with spatio-temporal full attention mechanism, positional encoding and the PredFormer encoder, and obtains the temporal features of the water level data; the feature fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the obtained spatio-temporal features of precipitation data of multiple meteorological stations and the temporal features of the water level data to obtain the fused features; the multi-layer perceptron outputs the water level prediction result based on the fused features;

[0191] The water level prediction module 205 is used to predict the water level of the hydrological station to be predicted by using the trained water level prediction model based on the dual-stream spatio-temporal network PredFormer.

[0192] Since the device introduced in the second embodiment of the present invention is the device used to implement the water level prediction method based on multi-modal fusion and the dual-stream spatio-temporal network PredFormer in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and deformation of the device, so it will not be elaborated here. Any device used in the method of the first embodiment of the present invention falls within the scope of protection of the present invention.

[0193] Embodiment Three

[0194] Based on the same inventive concept, please refer to Figure 5 The present invention also provides a computer-readable storage medium 300, on which a computer program 311 is stored. When the program is executed by a processor, it implements the method described in Embodiment One.

[0195] Since the computer-readable storage medium introduced in the third embodiment of the present invention is the computer-readable storage medium used to implement the water level prediction method based on multi-modal fusion and the dual-stream spatio-temporal network PredFormer in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and deformation of the computer-readable storage medium, so it will not be elaborated here. Any computer-readable storage medium used in the method of the first embodiment of the present invention falls within the scope of protection of the present invention.

[0196] Embodiment Four

[0197] The present invention also provides a computer device. Please refer to Figure 6 It includes a memory 401, a processor 402, and a computer program 403 stored on the memory and executable on the processor. When the processor executes the program, it implements the method described in Embodiment One.

[0198] Since the computer device introduced in the fourth embodiment of the present invention is the computer device used to implement the water level prediction method based on multi-modal fusion and the dual-stream spatio-temporal network PredFormer in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and deformation of the computer device, so it will not be elaborated here. Any computer device used in the method of the first embodiment of the present invention falls within the scope of protection of the present invention.

[0199] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0200] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0201] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and variations.

Claims

1. A water level prediction method based on multi-modal fusion and the dual-stream spatio-temporal network PredFormer, characterized in that, Including: Determine the corresponding meteorological station according to the hydrological station to be predicted; Collect the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station; Preprocess the collected water level monitoring data and precipitation monitoring data; Build a water level prediction model based on the dual-stream spatio-temporal network PredFormer, and use the preprocessed data to train the water level prediction model based on the dual-stream spatio-temporal network PredFormer. Among them, the water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch inputs embeddings, positional encodings and PredFormer encoders through the PredFormer model with spatio-temporal full attention mechanism, and sequentially establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations to obtain the spatio-temporal features of precipitation data of multiple meteorological stations. The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embeddings, positional encodings and PredFormer encoders of the PredFormer model with spatio-temporal full attention mechanism to obtain the temporal features of the water level data; Features The fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the spatio-temporal features of precipitation data of multiple meteorological stations and the temporal features of water level data obtained, and obtains the fused features; The multi-layer perceptron outputs the water level prediction result based on the fused features; Use the trained water level prediction model based on the dual-stream spatio-temporal network PredFormer to predict the water level of the hydrological station to be predicted.

2. The water level prediction method based on the PredFormer with multi-modal and spatio-temporal full attention mechanism as claimed in claim 1, wherein, Determine the corresponding meteorological station according to the hydrological station to be predicted, including: Based on the DEM, divide the river basin range where the hydrological station to be predicted is located, and select the meteorological station located upstream of the hydrological station and within the river basin range as the corresponding meteorological station.

3. The water level prediction method based on the PredFormer with multi-modal and spatio-temporal full attention mechanism according to claim 1, characterized in that Preprocess the collected water level monitoring data and precipitation monitoring data, including: Perform missing value, outlier and standardization processing on the collected water level monitoring data and precipitation monitoring data.

4. The water level prediction method based on the multi-modal and spatio-temporal full attention mechanism PredFormer according to claim 1, characterized in that, Before the first branch inputs embeddings, positional encodings and PredFormer encoders through the PredFormer model with spatio-temporal full attention mechanism, and sequentially establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations to obtain the spatio-temporal features of precipitation data of multiple meteorological stations, the method further includes: Use a graph convolutional neural network to model the spatial position relationship of multi-station precipitation data, construct the association information between stations, and obtain multi-station structured graph data.

5. The water level prediction method based on the multi-modal and spatio-temporal full attention mechanism PredFormer according to claim 4, wherein, The first branch inputs embeddings, positional encodings and PredFormer encoders through the PredFormer model with spatio-temporal full attention mechanism, and sequentially establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations to obtain the spatio-temporal features of precipitation data of multiple meteorological stations, including: The input embedding of the PredFormer model with spatio-temporal full attention mechanism takes the input multi-site structured graph data as multiple input sub-channels. The data of each sub-channel is flattened into a one-dimensional vector, and then the one-dimensional vector is projected into the hidden dimension through a linear layer to obtain the corresponding tensor; Position encoding is used to encode the spatial positions and time step sequences of each site in an absolute position encoding manner; Multiple stacked gated transformer blocks of the PredFormer encoder are combined with the multi-head self-attention mechanism to obtain the spatio-temporal features of precipitation data at multiple meteorological stations based on the position encoding; 6. The water level prediction method based on the multi-modal and spatio-temporal full attention mechanism PredFormer according to claim 1, characterized in that, The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embedding of the PredFormer model with spatio-temporal full attention mechanism, position encoding, and the PredFormer encoder, and obtains the temporal features of the water level data, including: The input embedding of the PredFormer model with spatio-temporal full attention mechanism processes the input single-site water level data to obtain the corresponding tensor; The single-site water level data is position-encoded by using absolute position encoding; Multiple stacked gated transformer blocks of the PredFormer encoder are combined with the multi-head self-attention mechanism to obtain the temporal features of the water level data based on the position encoding; 7. The water level prediction method based on the PredFormer with multi-modal and spatio-temporal full attention mechanism as claimed in claim 1, wherein, The output layer of the efficient low-rank tensor fusion method with modality-specific factors uses an MLP regression layer.

8. A water level prediction device based on the multi-modal and spatio-temporal full attention mechanism PredFormer, characterized in that, Including: A meteorological station determination module for determining the corresponding meteorological station according to the hydrological station to be predicted; A data collection module for collecting the water level monitoring data of the hydrological station to be predicted and the precipitation monitoring data of the corresponding meteorological station; A data preprocessing module for preprocessing the collected water level monitoring data and precipitation monitoring data; A model construction and training module for constructing a water level prediction model based on the dual-stream spatio-temporal network PredFormer and training the water level prediction model based on the dual-stream spatio-temporal network PredFormer by using the preprocessed data. Among them, the water level prediction model based on the dual-stream spatio-temporal network PredFormer includes a feature extraction module, a feature fusion module, and a multi-layer perceptron. The feature extraction module includes a two-branch structure. The first branch sequentially establishes the spatial relationship between meteorological stations and the temporal relationship within meteorological stations through the input of the PredFormer model with spatio-temporal full attention mechanism, position encoding, and the PredFormer encoder to obtain the spatio-temporal features of precipitation data at multiple meteorological stations. The second branch processes the temporal relationship of the water level data of the hydrological station to be predicted through the input embedding of the PredFormer model with spatio-temporal full attention mechanism, position encoding, and the PredFormer encoder to obtain the temporal features of the water level data; Feature The fusion module uses an efficient low-rank tensor fusion method with modality-specific factors to perform feature-level fusion on the obtained spatio-temporal features of precipitation data at multiple meteorological stations and the temporal features of the water level data to obtain the fused features; The multi-layer perceptron outputs the water level prediction result based on the fused features; The water level prediction module is used to predict the water level of the hydrological station to be predicted by using the trained water level prediction model based on the dual-stream spatio-temporal network PredFormer.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the water level prediction method based on the multi-modal and spatio-temporal full attention mechanism PredFormer as described in any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the water level prediction method based on the multi-modal and spatio-temporal full attention mechanism PredFormer as described in any one of claims 1 to 7.

Images