CN117094432A - Ocean thermal wave event prediction method based on interpretive deep learning - Google Patents

Ocean thermal wave event prediction method based on interpretive deep learning Download PDF

Info

Publication number
CN117094432A
CN117094432A CN202311023675.3A CN202311023675A CN117094432A CN 117094432 A CN117094432 A CN 117094432A CN 202311023675 A CN202311023675 A CN 202311023675A CN 117094432 A CN117094432 A CN 117094432A
Authority
CN
China
Prior art keywords
ocean
ocean thermal
prediction
deep learning
interpretive
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.)
Pending
Application number
CN202311023675.3A
Other languages
Chinese (zh)
Inventor
贺琪
宋巍
黄冬梅
杜艳玲
徐慧芳
朱姿杭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Ocean University
Original Assignee
Shanghai Ocean University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shanghai Ocean University filed Critical Shanghai Ocean University
Priority to CN202311023675.3A priority Critical patent/CN117094432A/en
Publication of CN117094432A publication Critical patent/CN117094432A/en
Pending legal-status Critical Current

Links

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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • 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
    • 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
    • G06Q10/08—Logistics, e.g. warehousing, loading or distribution; Inventory or stock management
    • G06Q10/083—Shipping
    • 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
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/02—Agriculture; Fishing; Mining

Abstract

The application provides a marine heat wave event prediction method based on interpretive deep learning, which comprises the following steps: selecting quantifiable marine environment elements, and establishing a prediction relationship between the target marine environment elements and the sea surface temperature based on an LSTM model; identifying ocean thermal wave events by combining a relative threshold method, and applying a desired gradient method to an LSTM model to quantify the time feature importance of input elements in ocean thermal wave prediction; and explaining the relation between the ocean thermal wave event and the ocean observation data from the feature importance scores, so as to realize the interpretation analysis of the ocean thermal wave event. The ocean thermal wave event prediction method based on the interpretable deep learning can accurately identify and predict the ocean thermal wave event and can explain the reason and mechanism of the prediction result.

Description

Ocean thermal wave event prediction method based on interpretive deep learning
Technical Field
The application relates to the technical field of ocean thermal wave discovery, in particular to an ocean thermal wave event prediction method based on interpretive deep learning.
Background
Ocean thermal waves are generally defined as a discrete, prolonged, abnormal warm water event occurring at a specific location that may have an impact on the marine ecosystem, fishery, shipping, travel industry, etc., and may also have an impact on human health, which is generally defined as a phenomenon in which the ocean surface temperature continues to be higher than an expected or historical average level for a certain period of time.
In the field of marine heat wave prediction, there are methods for predicting marine heat waves, which are generally based on models and algorithms in the fields of weather, oceanography, physics, etc., for example, a convolutional neural network-based method is used for predicting marine heat waves. However, in the deep learning method, the conventional neural network structure often has an overfitting problem, so that a prediction result is not accurate enough, and meanwhile, the interpretability is also lacking. In the sea surface temperature prediction task, due to the complexity of the sea environment, the traditional deep learning method is difficult to accurately predict and explain the change trend of the sea surface temperature, and meanwhile, the factors such as sea environment elements, climate change and the like are difficult to be integrated into a prediction model, so that the reliability and the interpretability of a prediction result are greatly influenced. In addition, the traditional deep learning method often needs a large amount of labeling data and computing resources, and the problems of fitting and the like easily occur in the training process, so that the wide application of the method in practical application is limited.
Therefore, a deep learning method with better interpretability is needed to predict ocean waves, the defects of the traditional method and the existing deep learning method can be overcome, the form and the characteristics of the ocean waves can be accurately identified and predicted, the reasons and the mechanisms of the prediction result can be interpreted, the prediction accuracy and the interpretability of the ocean waves are improved, and better services are provided for the fields of ocean ecosystems, fishery, shipping, travel industry and the like.
Disclosure of Invention
Aiming at the defects in the prior art, the application aims to provide the ocean thermal wave event prediction method based on the interpretable deep learning, which can accurately identify and predict the ocean thermal wave event and can explain the reason and mechanism of the prediction result.
In order to solve the problems, the technical scheme of the application is as follows:
an ocean thermal wave prediction method based on interpretive deep learning comprises the following steps:
selecting quantifiable marine environment elements, and establishing a prediction relationship between the target marine environment elements and the sea surface temperature based on an LSTM model;
identifying ocean thermal wave events by combining a relative threshold method, and applying a desired gradient method to an LSTM model to quantify the time feature importance of input elements in ocean thermal wave prediction;
and explaining the relation between the ocean thermal wave event and the ocean observation data from the feature importance scores, so as to realize the interpretation analysis of the ocean thermal wave event.
Preferably, the step of selecting quantifiable marine environmental elements and establishing a predictive relationship between the target marine environmental elements and the sea surface temperature based on the LSTM model specifically includes: and predicting the sea surface temperature by taking the air pressure and the air speed as input variables, and establishing a nonlinear prediction relation between the target sea area environment element and the sea surface temperature based on the LSTM model.
Preferably, the LSTM model comprises an LSTM layer and a fully connected layer, using as input two marine environment variables of air pressure and wind speed, and 180 time steps of historical data as input sequences to predict sea surface temperature.
Preferably, the LSTM network model comprises a cell state vector [ c ] that assumes long term memory of the network t ]And a hidden state as a nonlinear transition output of the cell stateVector [ h ] t ]At each time step t, the cyclic unit receives the unit state and hidden state from the previous unit, and the current input x t As input and calculate the current cell state c t And hidden state h t To supply the subsequent unit to use, the hidden state h of the last time step T Finally, mapping the full connection layer to a single neuron to realize sea surface temperature prediction.
Preferably, the LSTM network model is expressed as a mathematical formula:
f t ＝σ(W fx x t +W fh h t-1 +b f )
i t ＝σ(W ix x t +W ih h t-1 +b i )
o t ＝σ(W ox x t +W oh h t-1 +b o )
h t ＝o t ⊙tanh(c t )
wherein each of W and b represents an estimated learnable weight and bias term during training, and the LSTM model was calculated using a sigmod activation function σ (. Cndot.) and a hyperbolic tangent function tanh (. Cndot.), where @ represents the multiplication of the elements.
Preferably, the step of identifying a marine heat wave event in combination with a relative thresholding method and applying a desired gradient method to the LSTM model to quantify the temporal feature importance of the input element in marine heat wave prediction considers that a marine heat wave event has occurred when the sea water temperature exceeds a 90 percentile threshold based on long-term climate for at least 5 days, and to calculate this threshold, a daily sea water temperature value is used, and a weather threshold and average value for each calendar day in one year are calculated within an 11 day window centered on that day, and then smoothing is performed using a 31 day moving window, and the intensity of the marine heat wave event is defined as the difference between the temperature peak during heat waves and the weather temperature.
Preferably, the step of identifying ocean thermal wave events in combination with a relative thresholding method and applying a desired gradient method to the LSTM model to quantify the temporal feature importance of the input elements in ocean thermal wave prediction, the desired gradient method being formulated as:this formula represents EG score +.>Wherein x is i And x' i Represents the ith baseline and target inputs, respectively,>representing the local gradient of the interpolation point of the network f between the baseline and the target input, x '+α (x-x') represents the straight-line path of the baseline (α=0) to the target input (α=1), U (0, 1) represents a uniform distribution between 0 and 1, D is given one baseline distribution, and the final desired value is EG score +.>
Preferably, the step of interpreting the relationship between the ocean thermal wave event and the ocean observation data from the feature importance score to realize the interpretation analysis of the ocean thermal wave event specifically includes: to reveal possible implicit links between ocean thermal events and ocean observations, firstly, based on LSTM predictive model, air pressure and wind speed are used as input variables to predict sea surface temperature, and the data set is as follows: 3, dividing the ratio into a training set and a testing set, selecting root mean square error as an evaluation index, and performing error analysis on the training model on the testing set.
Compared with the prior art, the method provided by the application adopts a deep learning model, can adaptively learn the influence of factors such as marine environment and climate change on the sea surface temperature, realizes accurate identification and prediction of the marine heat wave event, and can provide a prediction result with strong interpretability in the prediction process, so that a user can deeply understand the form and the characteristic of the marine heat wave and understand the reason and the mechanism of the prediction result, thereby providing more comprehensive and accurate information for the user.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the detailed description of non-limiting embodiments, given with reference to the accompanying drawings in which:
FIG. 1 is a flow chart of a marine heat wave event prediction method based on interpretive deep learning provided by an embodiment of the application;
FIG. 2 is a flowchart illustrating the prediction of ocean thermal events according to an embodiment of the present application;
FIG. 3 is a schematic diagram of an LSTM model structure according to an embodiment of the present application;
FIG. 4 is a diagram of defining a marine thermal event according to an embodiment of the present application;
fig. 5 is a graph of SST prediction results of a ZLG site and a ZFD site according to an embodiment of the present application;
FIG. 6 is a sequence diagram of the characteristic importance of a ZLG station ocean thermal wave event provided by an embodiment of the application;
fig. 7 is a sequence diagram of the characteristic importance of ocean thermal wave events of the ZFD station according to an embodiment of the present application.
Detailed Description
The present application will be described in detail with reference to specific examples. The following examples will assist those skilled in the art in further understanding the present application, but are not intended to limit the application in any way. It should be noted that variations and modifications could be made by those skilled in the art without departing from the inventive concept. These are all within the scope of the present application.
The application combines the sequence pattern mining and the interpretable deep learning method and is used for mining the superposition influence pattern of the thermal wave event from the complex marine time sequence data, thereby revealing the association relation between the marine environment elements and the occurrence of the thermal wave event.
Specifically, the application provides a marine heat wave event prediction method based on interpretable deep learning, as shown in fig. 1 and 2, comprising the following steps:
s1: selecting quantifiable marine environment elements, and establishing a prediction relationship between the target marine environment elements and the sea surface temperature based on an LSTM model;
specifically, sea surface temperatures are predicted using marine environmental elements such as air pressure, wind speed, etc. as input variables, as they are directly related to the formation mechanism of ocean thermal wave events, and a nonlinear prediction relationship between the target sea area environmental elements and sea surface temperatures is established based on the LSTM model.
1. LSTM model
In this embodiment, LSTM networks are used as deep learning models, LSTM being one of the most popular deep learning architectures for modeling time series problems, with significant advantages in dealing with long time series data predictions of sea surface temperature in particular. Due to its cyclic structure and unique gating mechanism, it can effectively capture the non-linearity and time dependence between variables.
As shown in FIG. 3, the model used in this embodiment includes an LSTM layer and a fully connected layer, and uses two marine environment variables of air pressure and wind speed, 180 time steps in history as the input sequence [ x ] t ](wherein t is [1, T)]T=180) to predict sea surface temperature. The 180 time steps are chosen taking into account the periodic characteristics of the thermal events and the efficiency based on preliminary experiments, the LSTM network takes the form of a chain, consisting of repeated cyclic units, allowing the information to pass from one time step to the next. LSTM comprises a cell state vector [ c ] which bears the long-term memory of the network t ]And a hidden state vector h as a nonlinear transition output of cell state t ]. In each step t, the cyclic unit receives a signal from the previous unit (i.e., c t-1 And h t-1 ) Cell state and hidden state of (a) and current input x t As input and calculate the current cell statec t And hidden state h t To supply the subsequent units, the hidden state h of the last time step T Ultimately mapped to individual neurons, i.e. predicted sea surface temperatures, by one full connection.
First, a forget gate, a candidate cell state, an input gate and an output gate are calculated based on the previous hidden state, the current input and four multi-layer perceptrons, then the previous cell state c t-1 And carrying out linear updating according to the derived forgetting gate, the candidate unit state and the input gate. Given output gate and new cell state c t Obtaining a hidden state h through nonlinear transformation t . The mathematical formula for LSTM networks can be expressed as:
f t ＝σ(W fx x t +W fh h t-1 +b f )
i t ＝σ(W ix x t +W ih h t-1 +b i )
o t ＝σ(W ox x t +W oh h t-1 +b o )
h t ＝o t ⊙tanh(c t )
in this model, each W and b represents an estimated learnable weight and bias term during training. The model uses the sigmod activation function sigma (-) and the hyperbolic tangent function tanh (-) for calculation, and the ≡in the formula represents element multiplication. There are three gates in the model, with values between 0 and 1, which can control the reservation, addition and output of information streams. In this way, the model can accumulate information from previous and current time steps. Final hidden state h t The inputs of all time steps are considered, the proportion of which depends on the gate setting. The model may involveA plurality of hidden units such that c t And h t Becomes a vector of length 16. The full connection layer maps these 16 features to 1 feature using a learnable weight term W d And LSTM output h of last time step T And (3) performing calculation, wherein the formula is as follows: y=w d h T 。
S2: identifying ocean thermal wave events by combining a relative threshold method, and applying a desired gradient method to an LSTM model to quantify the time feature importance of input elements in ocean thermal wave prediction;
specifically, marine thermal wave events are identified in combination with a relative thresholding method, and a desired gradient method is applied to the trained LSTM model to quantify the temporal feature importance of input elements (i.e., barometric pressure, wind speed) in marine thermal wave prediction.
2. Ocean thermal wave event definition
Ocean thermal waves refer to abnormal warm water events occurring in the ocean and are characterized by a duration, intensity, rate of evolution, and spatial extent. In general, ocean waves are defined as discrete events at a particular location that last at least five consecutive extremely high temperature days with definite start and end dates. The extremely high ocean temperature means that the sea water temperature exceeds the 90 percentile threshold based on long-term climate and remains at this threshold T for at least five consecutive days threshold The above. To calculate this threshold, a daily sea water temperature value may be used, and the weather threshold and average for each calendar day of the year are calculated over an 11 day window centered on that day, and then smoothed using a 31 day moving window. In this definition, the seawater temperature is allowed to briefly fall below the threshold, but the time below the threshold cannot continue beyond 2 days. Intensity of ocean thermal wave event I MHW Is defined as the peak temperature and the climatic temperature T during the heat wave clim Is a difference between (a) and (b). The ocean thermal wave threshold is used in this embodiment to quantitatively analyze ocean thermal wave events and use predicted SST data to identify the time of occurrence of these events. A marine heat event, in particular, is considered to occur when the sea water temperature exceeds the 90 percentile threshold based on long term climate for at least 5 daysThe definition is shown in fig. 4.
Applying a desired gradient method to the trained LSTM model to obtain a characteristic importance score of the air pressure and wind speed before each heat wave event determined by the process line occurs;
3. expected gradient interpretation method
The purpose of the gradient (Expected Gradient, EG) method is to assign an importance score to each feature of a particular input, and the method of the present application assigns an importance score to the barometric pressure or wind speed, respectively, of each time step of the input sequence. A large positive number and a small negative number score indicate that the corresponding feature has a significant increasing or decreasing effect on the network output, while a importance score close to zero indicates that the corresponding feature has little effect on the network output.
Given a baseline profile D, by inputting all possible baselines, i.e. by density function p D' Weighting to obtain x' E D for gradient integration, and calculating EG score of ith featureThe formula is:
the equation contains two integrals, both of which can be regarded as desired values, and thus the above equation can be re-expressed as:
this formula is called the desired gradient, where U (0, 1) represents a uniform distribution between 0 and 1, in this embodiment, calculatedIs performed using the SHAP software package, which can provide various post hoc analyses for different neural networks. The application calculates each heat wave event which is determined by experiment>Obtained->The same dimensions as the corresponding input variables represent the time characteristic importance of air pressure and wind speed.
S3: and explaining the relation between the ocean thermal wave event and the ocean observation data from the feature importance scores, so as to realize the interpretation analysis of the ocean thermal wave event.
In this embodiment, the data of the national ocean science data center is adopted, and the data of the national station observation data comprises 14 stations including data of ocean weather, waves, temperature, salinity and the like. The embodiment is performed on two sites ZLG and ZFD, wherein the data time of the site ZLG is from 2015, 11, 1, to 2020, 8, 31, the data time of the ZFD is from 2012, 1, to 2013, 7, 12, the data of the two sites are daily average data, and the situation of missing data is estimated by adopting an interpolation method in consideration of the situation of missing data, so that the analysis effect of subsequent experiments caused by data missing is avoided.
In order to reveal possible hidden links between ocean thermal wave events and ocean observation data, firstly, based on an LSTM prediction model, the sea surface temperature is predicted by taking air pressure and wind speed as input variables, and for the robustness of an experiment, a data set is calculated according to 7:3 is divided into a training set and a testing set, root Mean Square Error (RMSE) is selected as an evaluation index, error analysis is carried out on the training model on the testing set, the prediction result is shown in fig. 5, and the error result is shown in table 1. Most RMSE is small (< 0.25), indicating that the network architecture effectively captures the potential dynamic relationship between the variables before the ocean thermal wave event occurs. In addition, five runs were performed for each site in the experiment, with a low standard deviation (< 0.05) for the five runs indicating the robustness of the model when trained with different segmented data sets.
TABLE 1
4. Results verification and analysis:
to verify the effectiveness of the method of the present application, the expected gradient method was applied to models trained by two sites in different segmentation datasets, extracting feature importance scores for 18 identified ocean thermal events. For different thermal wave events occurring at the same site, the scores extracted from the model generally exhibit similar patterns in terms of the relative importance of air pressure and wind speed for a particular time step, but differ slightly due to the randomness of model training. To explain the origin fromThe model of the thermal wave event mechanism, the experiment selects three representative thermal wave events from two sites, respectively, as shown in fig. 6. The thermal wave event occurring at ZLG site shows the air pressure +.>And the significance of the wind velocity characteristics are significant for an extended period of time before the event, which is affected by the superposition of atmospheric forces and the sea surface wind park irradiance, causes significant changes in wind speed and air pressure to initiate the ocean thermal wave event.
In contrast, fig. 7 shows the different importance patterns of ZFD site thermal wave events indicating that the effect of air pressure on the event is negligible compared to wind speed. Furthermore, only wind speeds occurring near the event are significantly exhibitedValues, while other recorded wind speed data have little apparent effect. Pattern analysis shows that the heat wave event may be caused by the fact that near-term sea surface wind field irradiation can also gather and enhance Ekman descending flow through ocean surface warm water, so that subsurface ocean anomaly continuous heating and subsurface ocean heat wave generation are caused. In this case, although the previous wind speeds in the site area may also be more or less heat intensiveThe extent of the wave event, but the nearest sea surface wind park irradiance has an absolute contribution to the occurrence of the event.
The foregoing describes specific embodiments of the present application. It is to be understood that the application is not limited to the particular embodiments described above, and that various changes or modifications may be made by those skilled in the art within the scope of the appended claims without affecting the spirit of the application. The embodiments of the application and the features of the embodiments may be combined with each other arbitrarily without conflict.

Claims (8)

1. An ocean thermal wave event prediction method based on interpretive deep learning is characterized by comprising the following steps:
selecting quantifiable marine environment elements, and establishing a prediction relationship between the target marine environment elements and the sea surface temperature based on an LSTM model;
identifying ocean thermal wave events by combining a relative threshold method, and applying an expected gradient method to an LSTM model to quantify the time feature importance of different input elements in ocean thermal wave prediction;
and explaining the relation between the ocean thermal wave event and the ocean observation data from the feature importance scores, so as to realize the interpretation analysis of the ocean thermal wave event.
2. The method for predicting ocean thermal events based on interpretive deep learning of claim 1, wherein the steps of selecting quantifiable ocean environmental elements and establishing a prediction relationship between target ocean environmental elements and ocean surface temperatures based on an LSTM model specifically comprise: and predicting the sea surface temperature by taking the air pressure and the air speed as input variables, and establishing a nonlinear prediction relation between the target sea area environment element and the sea surface temperature based on the LSTM model.
3. The method for predicting ocean thermal events based on interpretive deep learning of claim 2, wherein the LSTM model comprises an LSTM layer and a fully connected layer, and the sea surface temperature is predicted using two ocean environment variables of air pressure and wind speed as inputs and the history data of 180 time steps as an input sequence.
4. The method for predicting ocean thermal events based on interpretive deep learning of claim 3 wherein said LSTM network model comprises a cell state vector [ c ] assuming long term memory of the network t ]And a hidden state vector h as a nonlinear transition output of cell state t ]At each time step t, the cyclic unit receives the unit state and hidden state from the previous unit, and the current input x t As input and calculate the current cell state c t And hidden state h t To supply the subsequent unit to use, the hidden state h of the last time step T Finally, mapping the full connection layer to a single neuron to realize sea surface temperature prediction.
5. The method for predicting ocean thermal events based on interpretive deep learning of claim 4, wherein the LSTM network model is expressed as:
f t ＝σ(W fx x t +W fh h t-1 +b f )
i t ＝σ(W ix x t +W ih h t-1 +b i )
o t ＝σ(W ox x t +W oh h t-1 +b o )
h t ＝o t ⊙tanh(c t )
wherein each of W and b represents an estimated learnable weight and bias term during training, and the LSTM model was calculated using a sigmod activation function σ (. Cndot.) and a hyperbolic tangent function tanh (. Cndot.), where @ represents the multiplication of the elements.
6. The method for predicting ocean thermal events based on interpretive deep learning as claimed in claim 1, wherein the steps of identifying ocean thermal events in combination with a relative thresholding method and applying a desired gradient method to the LSTM model to quantify the temporal feature importance of the input elements in the ocean thermal event prediction, consider that an ocean thermal event occurs when the sea water temperature exceeds a 90 percentile threshold based on long-term climate for at least 5 days, use daily sea water temperature values for calculating the threshold, and calculate the weather threshold and average value for each calendar day in one year within an 11 day window centered on the day, and then use a 31 day moving window for smoothing, the intensity of the ocean thermal event being defined as the difference between the temperature peak and the weather temperature during thermal waves.
7. The method for predicting ocean thermal events based on interpretive deep learning of claim 1, wherein the step of identifying ocean thermal events in combination with a relative thresholding method and applying a desired gradient method to the LSTM model to quantify the temporal feature importance of the input elements in ocean thermal wave prediction, the desired gradient method is formulated as:this formula represents EG scoring for the ith featureWherein x is i And x' i Represents the ith baseline and target inputs, respectively,>office representing interpolation point of network f between baseline and target inputsPartial gradient, x '+α (x-x') represents the linear path of the baseline (α=0) to the target input (α=1), U (0, 1) represents a uniform distribution between 0 and 1, D is given a baseline distribution, and the final desired value is EG score +.>
8. The method for predicting ocean thermal events based on interpretive deep learning according to claim 1, wherein the step of interpreting the relationship between the ocean thermal events and the ocean observation data from the feature importance scores to realize the interpretive analysis of the ocean thermal events specifically comprises: to reveal possible implicit links between ocean thermal events and ocean observations, firstly, based on LSTM predictive model, air pressure and wind speed are used as input variables to predict sea surface temperature, and the data set is as follows: 3, dividing the ratio into a training set and a testing set, selecting root mean square error as an evaluation index, and performing error analysis on the training model on the testing set.
CN202311023675.3A 2023-08-15 2023-08-15 Ocean thermal wave event prediction method based on interpretive deep learning Pending CN117094432A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311023675.3A CN117094432A (en) 2023-08-15 2023-08-15 Ocean thermal wave event prediction method based on interpretive deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311023675.3A CN117094432A (en) 2023-08-15 2023-08-15 Ocean thermal wave event prediction method based on interpretive deep learning

Publications (1)

Publication Number Publication Date
CN117094432A true CN117094432A (en) 2023-11-21

Family

ID=88774434

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311023675.3A Pending CN117094432A (en) 2023-08-15 2023-08-15 Ocean thermal wave event prediction method based on interpretive deep learning

Country Status (1)

Country Link
CN (1) CN117094432A (en)

Similar Documents

Publication Publication Date Title
CN111899510B (en) Intelligent traffic system flow short-term prediction method and system based on divergent convolution and GAT
CN107909206B (en) PM2.5 prediction method based on deep structure recurrent neural network
CN110648014B (en) Regional wind power prediction method and system based on space-time quantile regression
CN110942194A (en) Wind power prediction error interval evaluation method based on TCN
CN109635245A (en) A kind of robust width learning system
CN111241755A (en) Power load prediction method
Dong et al. An integrated deep neural network approach for large-scale water quality time series prediction
CN112100911B (en) Solar radiation prediction method based on depth BILSTM
CN111814956A (en) Multi-task learning air quality prediction method based on multi-dimensional secondary feature extraction
CN111695290A (en) Short-term runoff intelligent forecasting hybrid model method suitable for variable environment
CN111858565B (en) Multi-dimensional high-precision track intelligent prediction method based on line segment clustering
Suryo et al. Improved time series prediction using LSTM neural network for smart agriculture application
CN110766060B (en) Time series similarity calculation method, system and medium based on deep learning
CN116702959A (en) Flood forecasting method based on distributed self-adaptive graph attention network
Lu et al. Neural network interpretability for forecasting of aggregated renewable generation
CN113780242A (en) Cross-scene underwater sound target classification method based on model transfer learning
CN114117852B (en) Regional heat load rolling prediction method based on finite difference working domain division
CN115512152A (en) Ship track classification method and system combining CNN (CNN) neural network and LSTM neural network
CN114596726B (en) Parking berth prediction method based on interpretable space-time attention mechanism
CN111753461A (en) Tidal water level correction method, target residual water level acquisition method, device and equipment
Jiang et al. Prediction of sea temperature using temporal convolutional network and LSTM-GRU network
CN117094432A (en) Ocean thermal wave event prediction method based on interpretive deep learning
CN115936062A (en) Photovoltaic power day-ahead interval prediction method based on scene generation and deep learning
Khan et al. Forecasting renewable energy for environmental resilience through computational intelligence
CN113053536B (en) Infectious disease prediction method, system and medium based on hidden Markov model

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