CN111680784B - Sea surface temperature deep learning prediction method based on space-time multidimensional influence - Google Patents
Sea surface temperature deep learning prediction method based on space-time multidimensional influence Download PDFInfo
- Publication number
- CN111680784B CN111680784B CN202010461315.1A CN202010461315A CN111680784B CN 111680784 B CN111680784 B CN 111680784B CN 202010461315 A CN202010461315 A CN 202010461315A CN 111680784 B CN111680784 B CN 111680784B
- Authority
- CN
- China
- Prior art keywords
- surface temperature
- sea surface
- space
- time
- deep learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000013135 deep learning Methods 0.000 title claims abstract description 42
- 238000012545 processing Methods 0.000 claims abstract description 25
- 230000000295 complement effect Effects 0.000 claims abstract description 15
- 238000005516 engineering process Methods 0.000 claims abstract description 12
- 239000013598 vector Substances 0.000 claims description 33
- 230000008569 process Effects 0.000 claims description 19
- 238000013507 mapping Methods 0.000 claims description 5
- 230000006870 function Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 description 9
- 102100022443 CXADR-like membrane protein Human genes 0.000 description 7
- 101000901723 Homo sapiens CXADR-like membrane protein Proteins 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000003442 weekly effect Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000001932 seasonal effect Effects 0.000 description 2
- 238000000547 structure data Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 241000288105 Grus Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 239000010931 gold Substances 0.000 description 1
- 229910052737 gold Inorganic materials 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000005855 radiation Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention provides a sea surface temperature deep learning prediction method based on space-time multidimensional influence. The method of the invention comprises three steps: 1. quantifying spatial impact: and quantifying the space influence force suffered by the target observation point by utilizing the sea surface temperature data of the neighbor observation points of the target observation point, and constructing a multidimensional space-time sea surface temperature data set on the target sea area. 2. Data complement processing: and carrying out data complement processing on the data set by using a neighbor data mean value method. 3. Establishing a prediction model: and combining deep learning technologies such as GRU, CNN and MLP, and the like, and establishing a sea surface temperature deep learning prediction model based on space-time multidimensional influence, namely a convolution gating circulation unit multi-layer perceptron (Convolutional GRU with Multilayer Perceptron, CGMP). The method integrates space-time multidimensional influence, combines a deep learning technology mode, establishes a sea surface temperature prediction model with high precision, and can be widely applied to sea surface temperature predictions of different sea areas and different scales.
Description
Technical Field
The invention relates to a sea surface temperature prediction method.
Background
Sea surface temperature (Sea Surface Temperature, SST), also known as sea surface temperature, refers to the temperature of water near the surface of the ocean. The exact meaning of the sea surface will vary depending on the measurement method used, but is typically between 1 mm and 20 meters below the sea surface. Sea surface temperature plays a fundamental role in the exchange of energy between the ocean and the atmosphere, and is also a very important parameter in ocean dynamics and climate change. The sea surface temperature can have profound effects on global climate, marine ecosystem and even marine life, and the well-known el-nino and lanina phenomena are caused by abnormal sea surface temperature (SST analysis, SSTA) changes. Climate models driven by observed or simulated sea surface temperatures or sea surface temperature anomalies are often used as a standard tool for seasonal climate prediction. Sea surface temperature is also often used as a key factor in the study of marine ecosystems and marine organisms. Therefore, the method has important significance in accurately and effectively observing and predicting the sea surface temperature. It not only can make human better understand global climate and marine ecosystem, but also can be an important component for marine related field application, such as extreme weather forecast, seasonal weather forecast, marine organism research and sustainable development fishery, etc. However, various physical and environmental factors (heat flux, radiation, and solar wind near the sea surface) affect changes in sea surface temperature, making it highly random and uncertain. Therefore, it is still a challenge to provide a sea surface temperature prediction method with high accuracy.
Sea surface temperature prediction methods can be categorized into two categories. One is a numerical method and the other is a data driving method. The conventional numerical method is not easy to model to describe the change in sea surface temperature because it requires much knowledge of the ocean and atmosphere domain and predicts sea surface temperature with relatively low resolution, typically on the ocean and even global scale. Compared to the numerical method, the data driving method, particularly the deep learning method, requires less knowledge of the ocean and atmosphere fields and can predict the sea surface temperature with high resolution on a smaller size, and therefore, it is easier to build a model to accurately predict the sea surface temperature of the target sea area. However, the sea surface temperature deep learning prediction method based on time influence generally improves the prediction performance by enhancing the capability of processing time information, and the method is often focused on capturing the time influence of the historical sea surface temperature, but ignores the integral space influence of the target sea area on the target observation point, so that the method has obvious performance bottleneck.
Disclosure of Invention
According to the sea surface temperature deep learning prediction model based on the space-time multidimensional influence, the performance bottleneck problem is solved by means of fusion of the space-time multidimensional influence and deep learning technologies such as a circulating neural network (Recurrent Neural Network, RNN), a convolutional neural network (Convolution Neural Network, CNN) and a multi-layer perceptron (Multilayer Perceptron, MLP), and the like, so that the accuracy of sea surface temperature prediction is further improved.
The invention is realized by the following technical scheme:
a sea surface temperature deep learning prediction method based on space-time multidimensional influence is characterized in that a sea surface temperature prediction model with high precision, namely a convolution gating circulation unit multi-layer perceptron (Convolutional GRU with Multilayer Perceptron, CGMP), is established by fusing the space-time multidimensional influence and combining a deep learning technology. The model can overcome the performance bottleneck of the sea surface temperature deep learning prediction model based on time influence, and can further improve the accuracy of sea surface temperature prediction. The method comprises the following specific steps:
A. quantifying spatial impact: quantifying the space influence force suffered by the neighbor observation point sea surface temperature data of the target observation point by utilizing the sea surface temperature data of the neighbor observation point, and constructing a multidimensional space-time sea surface temperature data set on the target sea area;
B. data complement processing: performing data complement processing on the data set by using a neighbor data average method;
C. establishing a prediction model: and establishing a sea surface temperature deep learning prediction model based on space-time multidimensional influence by combining deep learning technologies such as GRU, CNN, MLP and the like, namely a convolution gating circulation unit multi-layer perceptron CGMP.
And B, quantifying the space influence suffered by the sea surface temperature data of the neighbor observation point of the target observation point by utilizing the sea surface temperature data of the neighbor observation point of the step A, and constructing a multidimensional space-time sea surface temperature data set on the target sea area, wherein the specific steps are as follows:
a1: the sea surface may be divided into grids based on longitude, latitude, and resolution information. One grid is a sea surface temperature observation point. The sea surface temperature prediction task can be simply expressed as using historical sea surface temperature data at a target sea surface temperature observation point to predict future sea surface temperature data. In order to quantitatively quantify the space influence of the target sea surface temperature observation points, the space influence of the target observation points is assumed to be all from eight adjacent sea surface temperature observation points, namely, neighbor points, so that a multi-dimensional space-time sea surface temperature data set, namely, a sea surface temperature (daily average value/Zhou Junzhi/month average value) complete data set and a sea surface temperature (daily average value/Zhou Junzhi/month average value) ocean data set can be constructed on a target sea area, wherein the complete data set comprises data with space information loss, and the ocean data set excludes data with space information loss.
And B, performing data complement processing on the data set by using a neighbor data mean method, wherein the specific steps are as follows:
and B1, performing data complement processing on the complete data set by using a neighbor data average method. For some target observation points near the coast, some of their neighbor points may be located in a land area, and thus there is no sea surface temperature observation data on these neighbor points, which are called land sites. So that data completion processing is performed on land sites located in the land area for unified and standardized input of the prediction model. For land sites, the full sea surface temperature data is the average value of the real sea surface temperature observation data of eight neighbor points, namely a neighbor data average value method. The advantage of this data complement method is that it is practical and applicable to all different cases of absence. If other land points exist in the eight neighboring points, the average value of the rest sea points is calculated after the elimination.
And C, establishing a sea surface temperature deep learning prediction model based on space-time multidimensional influence by combining deep learning technologies such as GRU, CNN and MLP, namely a convolution gating circulation unit multi-layer perceptron CGMP, wherein the specific steps are as follows:
c1 and CGMP are end-to-end sea surface temperature prediction models, the input of the models is a historical SST sequence, and the output of the models is a predicted future SST sequence. The historical SST sequence comprises neighbor information and historical information of a target sea surface temperature observation point. The CGMP firstly processes the neighbor information of the target observation point in the space dimension through a convolution layer to obtain an intermediate vector, namely an empty SST sequence. The CGMP then processes the history of the spatiotemporal SST sequence in the time dimension by a GRU to obtain a hidden state vector. Finally, the multi-layer perceptron maps the hidden state vector onto the predicted result, i.e., the future SST sequence. The convolution layer is good at processing data similar to grid structure, and the GRU is good at processing sequence data, so that the combination collocation of the convolution layer and the GRU can fully and effectively mine neighbor and history information in the space-time dimension. Wherein:
c11, the history SST sequence X is a tensor with a dimension of kx3×3, and includes neighbor information (a matrix with a dimension of 3×3) and history information of k time units of the target sea surface temperature observation point. Because convolutional neural networks are good at processing grid-like structure data, CGMP first processes neighbor information of a target observation point in a spatial dimension through one convolutional layer. The depth of the convolution layer is 1, the size of the convolution kernel K is 3×3, the step size is 1, and the padding is 0. After the convolution operation is carried out on X, a space-time SST sequence S with the dimension of k multiplied by 1 is obtained, and the operation process is defined as follows:
S＝(s 1 ，s 2 ，...，s k )
s i ＝X i ★K+b s ，i∈{1，2，...，k}
wherein s is i Representing the processed spatio-temporal history information, +. k,1 –w k,9 Weights representing the convolution kernel K, b s Representing the corresponding bias.
C12, the space-time SST sequence S contains the historical information of k time units of the target sea surface temperature observation point. Since the GRU is good at processing the sequence data, the CGMP then processes the history information of the target observation point in the time dimension by one GRU. The GRU receives two inputs at each time step: one is the spatiotemporal history information S from the spatiotemporal SST sequence S i The other is the hidden state vector H from the previous time step (i-1) Or the initial hidden state vector H 0 . History information is processedThe GRU also generates a new hidden state vector H at each time step i . So the GRU will generate k hidden state vectors, H 1 –H k . The process is defined as follows:
r i ＝σ(W rs s i +W rH H (i-1) +b r )
z i ＝σ(W zs s i +W zH H (i-1) +b z )
n i ＝tanh(W ns s i +b ns +r i *(W nH H (i-1) +b nH ))
H i ＝(1-z i )*n i +z i *H (i-1)
wherein H is i Hidden state vector s representing GRU at time step i i Representing spatio-temporal history information at time step i, H (i-1) Represents the hidden state vector at time step i-1 or represents the initial hidden state vector at time step 0, r i 、z i And n i Representing the reset gate, update gate and new gate of the GRU, respectively. Sigma represents a sigmoid function, x represents a Hadamard product. W represents the corresponding weight and b represents the corresponding bias.
Finally, CGMP uses a multi-layer perceptron to hide the state vector H k Mapped onto the predicted result, i.e., future SST sequence Y. The mapping procedure is defined as follows:
Y＝f MLP (W Y H k +b Y )
where WY represents the corresponding weight and bY represents the corresponding bias.
The invention has the beneficial effects that:
1. the traditional sea surface temperature deep learning prediction method based on time influence is often focused on capturing the time influence of historical sea surface temperature, but ignores the integral space influence of a target sea area on a target observation point, so that the method has obvious performance bottleneck. The newly proposed sea surface temperature deep learning prediction method based on space-time multidimensional influence solves the performance bottleneck problem by fusing the space-time multidimensional influence and combining with the deep learning technologies such as RNN, CNN, MLP and the like, and further improves the accuracy of sea surface temperature prediction.
2. The method utilizes sea surface temperature data of a neighbor observation point of a target observation point to quantify the space influence force suffered by the target observation point. The method uses a neighbor data average method to carry out data complement processing on the space information.
3. By fusing space-time multidimensional influence and combining a deep learning technology, a sea surface temperature prediction model with high precision is established. The model can overcome the performance bottleneck of the sea surface temperature deep learning prediction model based on time influence, and can be widely applied to sea surface temperature predictions of different sea areas and different scales.
Drawings
FIG. 1 is a general flow chart of the sea surface temperature deep learning prediction method based on space-time multidimensional influence of the invention.
Fig. 2 is a schematic diagram of the principle of quantifying the spatial influence suffered by the neighboring observation point sea surface temperature data of the target observation point in the step a in fig. 1.
Fig. 3 is a schematic diagram of the data complement processing of the data set using the neighbor data mean method in the step B in fig. 1.
Fig. 4 is a schematic diagram of the model principle of the multi-layer perceptron CGMP of the convolution-gated loop unit described in step C of fig. 1.
Figure 5 shows the performance comparisons of CGMP, NGMP, CLMP and NLMP on a complete data set of sea surface temperature daily averages in Bohai.
Figure 6 shows a comparison of CGMP, NGMP, CLMP and NLMP performance over a complete data set of the sea surface temperature weekly averages of the Bohai sea.
FIG. 7 shows the performance comparisons of CGMP, FC-LSTM and GED on a complete sea surface temperature dataset of Bohai sea.
FIG. 8 shows the performance comparisons of CGMP, FC-LSTM and GED on the sea temperature marine dataset of Bohai sea.
FIG. 9 shows the performance comparisons of CGMP, FC-LSTM and GED on a complete data set of sea surface temperatures in the south China sea.
FIG. 10 shows the performance comparisons of CGMP, FC-LSTM and GED on a south sea surface temperature marine dataset.
Detailed Description
The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
The conception of the invention is as follows: firstly, quantifying the space influence suffered by a target sea surface temperature observation point, and constructing a multi-dimensional space-time sea surface temperature data set on a target sea area. The data set is then subjected to a data preprocessing and a data completion processing. And finally, establishing a sea surface temperature deep learning prediction model based on space-time multidimensional influence by combining RNN (LSTM or GRU), CNN, MLP and other deep learning technologies.
In this embodiment, the experiments were all completed on a Shanghai university machine learning platform with two Intel Xeon Gold 6130 CPUs, 192GB RAM, four Nvidia Tesla V100 GPUs and eighteen Nvidia Tesla P100 GPUs. In the aspect of establishing a model, the CGMP and the comparison model are realized based on an open-source deep learning tool PyTorch. In terms of experimental setup, adam was chosen as the optimization algorithm for the experiment, the batch size was set to 256 and the number of iterations was set to 1000. In terms of performance evaluation, experiments evaluate the sea surface temperature prediction performance of the model by calculating mean square error (Mean Square Error, MSE) and mean absolute error (Mean Absolute Error, MAE). The smaller the mean square error or mean absolute error obtained by the model, the better the predictive performance that represents it. The detailed definitions of MSE and MAE are as follows:
wherein Y represents a predicted future SST sequence, Y 'represents a corresponding observed SST sequence, Y represents a predicted future SST, Y' represents a corresponding true value, and l represents the magnitude of the prediction scale.
The sea surface temperature deep learning prediction method based on space-time multidimensional influence, as shown in fig. 1-4, comprises the following steps:
A. quantifying spatial impact: and quantifying the space influence force suffered by the target observation point by utilizing the sea surface temperature data of the neighbor observation points of the target observation point, and constructing a multidimensional space-time sea surface temperature data set on the target sea area. The method comprises the following specific steps:
a1, sea surface can be divided into grids according to longitude, latitude, and resolution information. One grid is a sea surface temperature observation point. The sea surface temperature prediction task can be simply expressed as using historical sea surface temperature data at a target sea surface temperature observation point to predict future sea surface temperature data. In order to quantitatively quantify the space influence of the target sea surface temperature observation points, the space influence of the target observation points is assumed to be all from eight sea surface temperature observation points adjacent to the target observation points and called neighbor points, so that a multidimensional space-time sea surface temperature data set, namely a Bohai sea surface temperature (daily average value/Zhou Junzhi/month average value) complete data set, a Bohai sea surface temperature (daily average value/Zhou Junzhi/month average value) ocean data set, a south sea surface temperature (daily average value/Zhou Junzhi/month average value) complete data set and a south sea surface temperature (daily average value/Zhou Junzhi/month average value) ocean data set can be constructed on Bohai sea and south sea areas. The complete data set contains data with missing spatial information, and the marine data set excludes data with missing spatial information.
B. Data complement processing: and carrying out data complement processing on the data set by using a neighbor data mean value method. The method comprises the following specific steps:
and B1, performing data complement processing on the complete data set by using a neighbor data average method. For some target observation points near the coast, some of their neighbor points may be located in a land area, and thus there is no sea surface temperature observation data on these neighbor points, which are called land sites. Therefore, in order to unify and standardize the input of the prediction model, the chapter performs data completion processing on land sites located in land areas. For land sites, the full sea surface temperature data is the average value of the real sea surface temperature observation data of eight neighbor points, namely a neighbor data average value method. The advantage of this data complement method is that it is practical and applicable to all different cases of absence.
C. Establishing a prediction model: and establishing a sea surface temperature deep learning prediction model based on space-time multidimensional influence by combining deep learning technologies such as GRU, CNN, MLP and the like, namely a convolution gating circulation unit multi-layer perceptron CGMP. The method comprises the following specific steps:
c1 and CGMP are end-to-end sea surface temperature prediction models, the input of the models is a historical SST sequence, and the output of the models is a predicted future SST sequence. The historical SST sequence comprises neighbor information and historical information of a target sea surface temperature observation point. The CGMP firstly processes the neighbor information of the target observation point in the space dimension through a convolution layer to obtain an intermediate vector, namely an empty SST sequence. The CGMP then processes the history of the spatiotemporal SST sequence in the time dimension by a GRU to obtain a hidden state vector. Finally, the multi-layer perceptron maps the hidden state vector onto the predicted result, i.e., the future SST sequence. The convolution layer is good at processing data similar to grid structure, and the GRU is good at processing sequence data, so that the combination collocation of the convolution layer and the GRU can fully and effectively mine neighbor and history information in the space-time dimension. Wherein:
c11, the history SST sequence X is a tensor with a dimension of kx3×3, and includes neighbor information (a matrix with a dimension of 3×3) and history information of k time units of the target sea surface temperature observation point. Because convolutional neural networks are good at processing grid-like structure data, CGMP first processes neighbor information of a target observation point in a spatial dimension through one convolutional layer. The depth of the convolution layer is 1, the size of the convolution kernel K is 3×3, the step size is 1, and the padding is 0. After the convolution operation is carried out on X, a space-time SST sequence S with the dimension of k multiplied by 1 is obtained, and the operation process is defined as follows:
S＝(s 1 ，s 2 ，...，s k )
s i ＝X i ★K+b s ，i∈{1，2，...，k}
wherein s is i Representing the processed spatio-temporal history information, +. k,1 –w k,9 Weights representing the convolution kernel K, b s Representing the corresponding bias.
C12, the space-time SST sequence S contains the historical information of k time units of the target sea surface temperature observation point. Since the GRU is good at processing the sequence data, the CGMP then processes the history information of the target observation point in the time dimension by one GRU. The GRU receives two inputs at each time step: one is the spatiotemporal history information S from the spatiotemporal SST sequence S i The other is the hidden state vector H from the previous time step (i-1) Or the initial hidden state vector H 0 . After the history information is processed, the GRU also generates a new hidden state vector H at each time step i . So the GRU will generate k hidden state vectors, H 1 –H k . The process is defined as follows:
r i ＝σ(W rs s i +W rH H (i-1) +b r )
z i ＝σ(W zs s i +W zH H (i-1) +b z )
n i ＝tanh(W ns s i +b ns +r i *(W nH H (i-1) +b nH ))
H i ＝(1-z i )*n i +z i *H (i-1)
wherein H is i Hidden state vector s representing GRU at time step i i Representing spatio-temporal history information at time step i, H (i-1) Represents the hidden state vector at time step i-1 or represents the initial hidden state vector at time step 0, r i 、z i And n i Representing the reset gate, update gate and new gate of the GRU, respectively. Sigma represents a sigmoid function, x represents a Hadamard product. W represents the corresponding weight and b represents the corresponding bias.
Finally, CGMP uses a multi-layer perceptron to hide the state vector H k Mapped onto the predicted result, i.e., future SST sequence Y.
The mapping procedure is defined as follows:
Y＝f MLP (W Y H k +b Y )
wherein W is Y Representing the corresponding weights, b Y Representing the corresponding bias.
Referring to fig. 5 and 6, performance comparisons of CGMP, NGMP, CLMP and NLMP on the Bohai sea surface temperature daily mean and Zhou Junzhi complete data set are shown. In the experiment, the convolution layer and the full connection layer are subjected to comparison of spatial information extraction performance, and the GRU and the LSTM are subjected to comparison of time information extraction performance. Therefore, besides CGMP, there are three other corresponding comparison methods for experiments, namely a Neighbor Full-connection gating cycle unit multi-layer sensor (Neighbor Full-Connected GRU with Multilayer Perceptron, NGMP), a convolution long-short-period memory multi-layer sensor (Convolutional LSTM with Multilayer Perceptron, CLMP) and a Neighbor Full-connection long-period memory multi-layer sensor (Neighbor Full-Connected LSTM with Multilayer Perceptron, NLMP). Meanwhile, the experiment also discusses the influence of the depth of the multi-layer perceptron on the prediction performance. As can be seen from comparing the experimental results of CGMP and NGMP, under the condition of the same depth of the multi-layer perceptron, the MSE and MAE of the CGMP are smaller than those of the NGMP, that is to say, the prediction performance of the CGMP on the complete data set of the average daily value of the sea surface temperature of the Bohai sea is better than that of the NGMP. This also indirectly illustrates that the ability of the convolution layer to capture spatial influence in the spatial dimension is superior to the fully connected layer. Comparing the experimental results of CGMP and CLMP, it can be found that under the same depth of the multi-layer perceptron, the MSE and MAE of CGMP are smaller than those of CLMP, that is to say, the prediction performance of CGMP on the complete data set of sea surface temperature daily average value of Bohai sea is better than that of CLMP. This also indirectly illustrates that the ability of the GRU to process history information in the time dimension is superior to LSTM. Comparing the experimental results of these four methods together, it can be found that CGMP using both the convolution layer and the GRU exhibits optimal predictive performance at a depth of 3 for the multi-layer perceptron. In summary, convolutional layers may capture spatial impact in the spatial dimension more efficiently than fully connected layers, and GRUs may process history information in the temporal dimension more efficiently than LSTM.
Referring to FIGS. 7 and 8, performance comparisons of CGMP, FC-LSTM and GED on Bohai sea surface temperature integrity and marine data sets are shown. As is evident from the table, the predictive performance of CGMP remains optimal in both daily, weekly and monthly average sea surface temperature predictions as compared to FC-LSTM and GED
Referring to FIGS. 9 and 10, there is shown the performance comparisons of CGMP, FC-LSTM and GED on the south sea surface temperature whole-number and sea dataset. It is evident from the table that the predictive performance of CGMP remains optimal in both daily, weekly and monthly average sea surface temperature predictions, as compared to FC-LSTM and GED.
In summary, according to the sea surface temperature deep learning prediction method based on the space-time multidimensional influence, a sea surface temperature prediction model with high precision is established by fusing the space-time multidimensional influence and combining a deep learning technology. The model can overcome the performance bottleneck of the sea surface temperature deep learning prediction model based on time influence, and can be widely applied to sea surface temperature predictions of different sea areas and different scales.
The description set forth herein with reference to the drawings and the detailed description is only intended to assist in understanding the methods and core concepts of the invention. The method according to the invention is not limited to the examples described in the specific embodiments, but other embodiments according to the method and idea according to the invention are also within the technical scope of the invention. The description should not be taken as limiting the invention.
Claims (5)
1. A sea surface temperature deep learning prediction method based on space-time multidimensional influence is characterized by comprising the following steps:
A. quantifying spatial impact: quantifying the space influence force suffered by the neighbor observation point sea surface temperature data of the target observation point by utilizing the sea surface temperature data of the neighbor observation point, and constructing a multidimensional space-time sea surface temperature data set on the target sea area;
B. data complement processing: performing data complement processing on the data set by using a neighbor data average method;
C. establishing a prediction model: by combining with deep learning technologies such as convolution operation, a gating circulation unit, a multi-layer perceptron and the like, a sea surface temperature deep learning prediction model based on space-time multidimensional influence is established, namely the convolution gating circulation unit multi-layer perceptron is called CGMP for short, the CGMP is an end-to-end sea surface temperature prediction model, the input of the model is a historical SST sequence,
the output of the model is a predicted future SST sequence, wherein the historical SST sequence comprises historical sea surface temperature data of a target sea surface temperature observation point and a neighbor point, CGMP firstly processes the neighbor information of the target observation point in a space dimension through a convolution layer to obtain an intermediate vector, an immediate empty SST sequence, and then processes the historical information of the immediate empty SST sequence in a time dimension through a gating circulation unit to obtain a hidden state vector, finally,
mapping the hidden state vector to a prediction result, namely a future SST sequence by using a multi-layer perceptron; the history SST sequence X is a tensor with a dimension of k multiplied by 3, comprises neighbor information and history data information of k time units of the target sea surface temperature observation point, the CGMP firstly processes the neighbor information of the target observation point in the space dimension through a convolution layer,
the depth of the convolution layer is 1, the size of the convolution kernel K is 3 multiplied by 3, the step length is 1, the filling is 0, and after convolution operation, the space-time SST sequence S with the dimension of K multiplied by 1 is obtained, and the operation process is defined as follows:
S(s1,S2,..,S)
s i ＝X i ★K+b s ，i∈{1，2，...，k}
wherein s is i Representing the processed spatio-temporal history information, +. k，1 –w k，9 Weights representing the convolution kernel K, b s Representing a phaseAnd a corresponding bias.
2. The sea surface temperature deep learning prediction method based on space-time multidimensional influence as claimed in claim 1, wherein the sea surface temperature deep learning prediction method is characterized in that: in step a, the sea surface is divided into grids according to longitude, latitude and resolution information, one grid is one sea surface temperature observation point, eight sea surface temperature observation points adjacent to the target observation point are called neighbor points, and a multi-dimensional space-time sea surface temperature dataset of the nine points can be constructed on the target sea area.
3. The sea surface temperature deep learning prediction method based on space-time multidimensional influence as claimed in claim 2, wherein the method is characterized by comprising the following steps of: in the step B, the data of the missing points in the nine points are complemented, the average value of the real sea surface temperature observation data of 8 neighbor points around the missing points is used as the missing point, and if the neighbor points of the missing points also have data missing, the missing points are eliminated in calculation.
4. The sea surface temperature deep learning prediction method based on space-time multidimensional influence as recited in claim 3, wherein the sea surface temperature deep learning prediction method is characterized by: the gated loop unit is abbreviated as GRU, which receives two inputs at each time step: one is the spatiotemporal history information S from the spatiotemporal SST sequence S i The other is the hidden state vector H from the previous time step (i-1) Or the initial hidden state vector H 0 After the history information is processed, the GRU also generates a new hidden state vector H at each time step i GRU generates k hidden state vectors, namely H 1 –H k The process is defined as follows:
r i ＝σ(W rs s i +W rH H (i-1) +b r )
z i ＝σ(W ss s i +W zH H (i-1) +b z )
n i ＝tanh(W ns s i +b ns +r i *(W nH H (i-1) +b nH ))
H i ＝(1-z i )*n i +z i *H (i-1)
wherein H is i Represents hidden state vector of GRU at time step i, si represents time-space history information at time step i, H (i-1) Represents the hidden state vector at time step i-1 or represents the initial hidden state vector at time step 0, r i 、z i And n i Respectively representing a reset gate, an update gate and a new gate of the GRU, sigma represents a sigmoid function, x represents Hadamard products, W represents corresponding weights, and b represents corresponding offsets.
5. The sea surface temperature deep learning prediction method based on space-time multidimensional influence as recited in claim 4, wherein the sea surface temperature deep learning prediction method is characterized by comprising the following steps of: CGMP uses a multi-layer perceptron to hide the state vector H k Mapping to the predicted result, i.e. future SST sequence Y, the mapping procedure is defined as follows:
Y＝f MLP (W Y H k +b Y )
wherein W is Y Representing the corresponding weights, b Y Representing the corresponding bias.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010461315.1A CN111680784B (en) | 2020-05-27 | 2020-05-27 | Sea surface temperature deep learning prediction method based on space-time multidimensional influence |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010461315.1A CN111680784B (en) | 2020-05-27 | 2020-05-27 | Sea surface temperature deep learning prediction method based on space-time multidimensional influence |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111680784A CN111680784A (en) | 2020-09-18 |
| CN111680784B true CN111680784B (en) | 2023-10-24 |
Family
ID=72434400
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010461315.1A Active CN111680784B (en) | 2020-05-27 | 2020-05-27 | Sea surface temperature deep learning prediction method based on space-time multidimensional influence |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111680784B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112465203A (en) * | 2020-11-19 | 2021-03-09 | 中国石油大学(华东) | Sea level height intelligent prediction and forecast system based on gate control circulation unit neural network, computer equipment and storage medium |
| CN117633712B (en) * | 2024-01-24 | 2024-04-19 | 国家卫星海洋应用中心 | Sea level height data fusion method, device and equipment based on multi-source data |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685767A (en) * | 2018-11-26 | 2019-04-26 | 西北工业大学 | A kind of bimodal brain tumor MRI dividing method based on Cluster-Fusion algorithm |
| CN109754605A (en) * | 2019-02-27 | 2019-05-14 | 中南大学 | A kind of traffic forecast method based on attention temporal diagram convolutional network |
| CN110176309A (en) * | 2019-05-28 | 2019-08-27 | 上海大学 | It is a kind of for predicting the medical data processing method of cardiovascular disease |
| CN110321603A (en) * | 2019-06-18 | 2019-10-11 | 大连理工大学 | A kind of depth calculation model for Fault Diagnosis of Aircraft Engine Gas Path |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10990874B2 (en) * | 2017-05-22 | 2021-04-27 | Sap Se | Predicting wildfires on the basis of biophysical indicators and spatiotemporal properties using a convolutional neural network |
| US11275989B2 (en) * | 2017-05-22 | 2022-03-15 | Sap Se | Predicting wildfires on the basis of biophysical indicators and spatiotemporal properties using a long short term memory network |
-
2020
- 2020-05-27 CN CN202010461315.1A patent/CN111680784B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685767A (en) * | 2018-11-26 | 2019-04-26 | 西北工业大学 | A kind of bimodal brain tumor MRI dividing method based on Cluster-Fusion algorithm |
| CN109754605A (en) * | 2019-02-27 | 2019-05-14 | 中南大学 | A kind of traffic forecast method based on attention temporal diagram convolutional network |
| CN110176309A (en) * | 2019-05-28 | 2019-08-27 | 上海大学 | It is a kind of for predicting the medical data processing method of cardiovascular disease |
| CN110321603A (en) * | 2019-06-18 | 2019-10-11 | 大连理工大学 | A kind of depth calculation model for Fault Diagnosis of Aircraft Engine Gas Path |
Non-Patent Citations (3)
| Title |
|---|
| Changjiang Xiao.A spatiotemporal deep learning model for sea surface temperature field prediction using time-series satellite data.《Environmental Modelling & Software》.2019,1-5页. * |
| Yuting Yang等.A CFCC-LSTM Model for Sea Surface Temperature Prediction.《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》.2018,全文. * |
| 姜英超.基于深度学习的海表温度遥感反演模型.《中国优秀博硕士学位论文全文数据库(硕士) 基础科学辑》.全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111680784A (en) | 2020-09-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Automatic concrete crack segmentation model based on transformer | |
| Braakmann-Folgmann et al. | Sea level anomaly prediction using recurrent neural networks | |
| CN111680784B (en) | Sea surface temperature deep learning prediction method based on space-time multidimensional influence | |
| CN114067019A (en) | Urban waterlogging risk map rapid prefabricating method coupling deep learning and numerical simulation | |
| Jin et al. | Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering | |
| CN115712873A (en) | Photovoltaic grid-connected operation abnormity detection system and method based on data analysis and infrared image information fusion | |
| CN114399073A (en) | Ocean surface temperature field prediction method based on deep learning | |
| CN114757904A (en) | Surface defect detection method based on AI deep learning algorithm | |
| Chen et al. | WSN sampling optimization for signal reconstruction using spatiotemporal autoencoder | |
| CN115933010A (en) | Radar echo extrapolation near weather prediction method | |
| Abubakr et al. | Application of deep learning in damage classification of reinforced concrete bridges | |
| Cao et al. | A photovoltaic surface defect detection method for building based on deep learning | |
| Chen et al. | Dynamic multiscale fusion generative adversarial network for radar image extrapolation | |
| Baydaroğlu et al. | Temporal and spatial satellite data augmentation for deep learning-based rainfall nowcasting | |
| CN117422695A (en) | CR-deep-based anomaly detection method | |
| CN115217152A (en) | Method and device for predicting opening and closing deformation of immersed tunnel pipe joint | |
| CN114676887A (en) | River water quality prediction method based on graph convolution STG-LSTM | |
| Hu et al. | Bayesian averaging-enabled transfer learning method for probabilistic wind power forecasting of newly built wind farms | |
| CN113962432A (en) | Wind power prediction method and system integrating three-dimensional convolution and light-weight convolution threshold unit | |
| Fan et al. | Urban digital twins for intelligent road inspection | |
| CN116912675B (en) | Underwater target detection method and system based on feature migration | |
| Naveen Venkatesh et al. | Photovoltaic Module Fault Detection Based on Deep Learning Using Cloud Computing | |
| CN117237781B (en) | Attention mechanism-based double-element fusion space-time prediction method | |
| CN117437204A (en) | Insulator defect rapid detection method and system based on YOLO-MID | |
| Xia et al. | A fault detection method for AUV based on multi-scale spatiotemporal feature fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |