Construction of Sea Surface Temperature Forecasting Model for Bohai Sea and Yellow Sea Coastal Stations Based on Long Short-Time Memory Neural Network

Abstract

In order to address the issue of large errors in predicting SST along the coast using numerical models, this study adopts LSTM, a deep learning method, to develop optimal SST prediction models. The Xiaomaidao Station is selected as an example, and then the method is then extended to 14 coastal stations along the Bohai Sea and the Yellow Sea. The results show that the SST prediction model based on LSTM effectively improves forecast accuracy. The mean absolute errors for 1–3-day SST forecasts of the optimal model at Xiaomaidao Station are 0.20 °C, 0.27 °C, and 0.31 °C, and the root mean square errors are 0.28 °C, 0.36 °C, and 0.41 °C, respectively, representing an average reduction of 78% compared to those of the numerical model. Extending this approach to other forecasting sites along the Bohai Sea and the Yellow Sea results in an average 61% reduction in forecast error when compared with the numerical model. Furthermore, it is found that utilizing an LSTM model can significantly save computational resources and improve the forecasting efficiency.

1. Introduction

The Bohai Sea and the Yellow Sea, located along the northern coastline of China, are abundant in marine resources such as fisheries, harbors, petroleum, and tourism. They have been one of the earliest areas in China to be developed and utilized for their marine resources, playing a crucial role in local economic development [1,2,3,4]. However, changes in the marine environment can significantly impact the sustainable development of the marine economy through alterations in ocean heat conditions, dynamic processes, and ecological environments [5,6,7,8]. Therefore, it is of great significance to study changes in the offshore marine environment.

Sea surface temperature (SST) is a fundamental and crucial element of the ocean. Abnormal changes in SST can result in variations in ocean circulation patterns, fluctuations in sea levels, and changes in the ecological environment [9,10,11,12,13,14], and even lead to extreme climate events such as extensive sea ice generation or marine heat waves [15,16,17]. For instance, at the beginning of 2010, SST in the Bohai Sea was unusually low, leading to early and rapid development of sea ice, causing significant impact on the region. The sea ice affected 61,000 people along the Bohai coast, damaged 7157 ships, froze 296 ports and docks along the coast, and damaged 20,787,000 hectares of aquaculture. Additionally, sea ice blocked 13 offshore islands, leaving residents unable to secure daily necessities and emergency supplies. According to statistics, the direct economic loss caused by sea ice in that year reached CNY 6.318 billion [15,16,17]. Another instance is an unprecedented marine heat wave event in August 2016 in the East China Sea where average SST exceeded 28.7 °C—significantly higher than the climate average by 1.8 °C. The heat wave had a significant impact on marine fisheries and aquaculture. For example, approximately 950,000 mu of sea cucumber aquaculture areas along Liaoning’s coast suffered economic losses totaling CNY 6.87 billion. Furthermore, the increased SST led to delayed seeding of wakame in Dalian and other coastal areas as well as dislodging a large number of seedlings from culture ropes, resulting in significant economic losses [18,19,20]. Therefore, understanding SST development trends and making timely accurate forecasts can provide necessary information for relevant departments to perform disaster prevention work, in order to reduce impacts caused by marine disasters effectively [18,19,20].

Currently, the operational prediction of SST mainly relies on two methods: numerical models and manual experience. Numerical model prediction has the advantage of including physical processes in the model, allowing for the simultaneous calculation of prediction results across the entire spatial field using large computers. The accuracy of numerical simulation prediction results is high in the vast sea area, but it is lower in coastal sea area due to factors such as local topography, boundary conditions, initial fields, and ocean currents. In contrast to numerical models, manual experience is more effective for coastal forecasting but requires more time and may result in subjective differences depending on forecasters.

In recent years, with the emergence of artificial intelligence (AI), deep learning has once again garnered attention. AI research fields primarily include intelligent robots, machine vision, image recognition, language recognition, natural language processing, and expert systems. The concept of deep learning was first proposed by Hinton et al. from the University of Toronto in 2006 [21], referring to the process of obtaining a deep network structure containing multiple levels based on sample data through specific training methods. Typical network structures used in deep learning include convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), and deep belief networks (DBNs). Among these structures, RNN is particularly useful for modeling sequence data where current output depends on previous outputs, which is mainly used for dealing with time series structures.

Long short-time memory (LSTM) is further developed on the basis of RNN by not only retaining its advantages but also addressing issues such as gradient disappearance or the explosion and lack of long-term memory. LSTM’s ability for long-term learning makes it suitable for solving predictive problems [22]. Currently, the LSTM method has been preliminarily applied in ocean forecasting [23,24,25]. For instance, Gao Libin et al. established a wave height prediction model using the LSTM method [26]. The MAE reached a minimum of 0.008 m, the RMSE reached a minimum of 0.012 m, and the correlation coefficient R reached a maximum of 0.999, indicating that LSTM has a good effect in wave height prediction. Gao Song et al. utilized LSTM to forecast ocean waves and compared them with numerical model results [27], and the RMSE and MAE decreased by 18% and 22%, respectively. Zhu Guizhong et al. adopt the LSTM-RNN method to predict the monthly mean SST of the following month in the Western Pacific Ocean, achieving an MAE of 0.15 °C and RMSE of 0.19 °C, significantly improving the accuracy of existing SST prediction models [28].

In this paper, the LSTM method is utilized to replace the numerical forecast model to build the SST intelligent forecast model in the coast of the Bohai Sea and the Yellow Sea, based on the operational SST forecast requirements. The goal is to enhance the prediction accuracy of numerical models and achieve a level comparable to manual empirical prediction. Firstly, an intelligent forecasting model is constructed using the Xiaomaidao Ocean Station as a case study, with the evaluation of forecasting error for the optimal intelligent model. Subsequently, this method is extended to 14 ocean stations along the Bohai Sea and Yellow Sea to construct forecasting models and evaluate their forecasting effects. Finally, limitations of current methods are discussed, and future work prospects are considered.

2. Data and Methods

2.1. SST Observation Data

The SST data use hour-by-hour observations from 14 ocean stations along the coast of Bohai Sea and Yellow Sea from 1 August 2018 to 31 July 2021. The observational data are mainly used for constructing and testing intelligent forecasting models of SST.

In this paper, Xiaomaidao Ocean Station is taken as an example to demonstrate the building process of the intelligent SST prediction model. Built in July 1959, the Xiaomaidao Ocean Station is situated in Xiaomaidao, Laoshan District, Qingdao, China (Figure 1). It stands out as one of the few marine environmental monitoring stations with comprehensive observation and monitoring projects in China. Additionally, it is among the earliest national demonstration stations to implement automated ocean observation, and the location of the measuring point has remained unchanged since the station was built, and the surrounding environment has not changed significantly. Surrounded by the sea and connected to the land by a seawall, Xiaomaidao has a park on the island but no permanent residents. Therefore, the observation and monitoring data collected at this station are very representative and can effectively reflect the fundamental characteristics and changing patterns of the marine environment off Qingdao.

2.2. Meteorological Forecast Data

The meteorological forecast data are sourced from the operational weather forecast system of North China Sea Marian Forecast and Hazard Mitigation Service. The system is based on the mesoscale meteorological model WRF, incorporating advanced three-dimensional variable data assimilation technology to form an atmospheric initial field to drive the regional atmospheric model. The data utilized for assimilation include conventional meteorological observation data such as GTS, buoys, and ocean stations, as well as non-conventional observation data such as satellites and aircraft. Then, combined with the parameterization scheme for weather forecasting in the Bohai Sea and the Yellow Sea, the operational model of a meteorological numerical forecast is formed, and refined numerical forecast products of meteorological elements for these regions are provided. The model has a maximum horizontal spatial resolution of 3 km, a time resolution of 1 h, and a running time of about 2 h. It can provide hourly weather forecast data for the next 7 days.

From the results of this model, hourly meteorological element data at ocean station locations were extracted including air temperature at 2 m above sea level, relative humidity at 2 m above sea level, wind speed at 10 m above sea level, wind direction at 10 m above sea level, surface heat flux, latent heat flux, etc. The data period is consistent with SST observations and covers 1 August 2018 to 31 July 2021.

2.3. SST Forecast Data

The results of SST numerical prediction are utilized to compare and verify the effect of the intelligent SST model. These predictions come from a three-dimensional temperature–salt–flow regional ocean modeling system (ROMS) operated by North China Sea Marian Forecast and Hazard Mitigation Service.

Three regional ocean models are established using multiple nesting techniques. The large area covers the entire Northwest Pacific Ocean (99°–148° E, 9° S–44° N), with a horizontal resolution of 0.1° and 25 vertical layers. The central area is the East China Sea (117°30′–135° E, 24°–41° N), with a horizontal resolution of 1/30° and 16 vertical layers. The small area covers the Yellow and Bohai Sea area (117°30′–128° E, 32°–41° N), with a horizontal resolution of 1/60° and 16 vertical layers. The output results from the Global ocean model (HYCOM + NCODA Global Analysis) are used as initial and boundary value fields for the large-area model. The simulated values from the upper-level region are used as initial and boundary value fields for both medium- and small-region models. The operation time of the models is about 0.5 h, which can provide hourly SST forecast data for the next 7 days. The construction and operation process of the model, as well as the stability test, are detailed in references [9,29,30,31]. The data used spans from 1 August 2020 to 31 July 2021.

2.4. Data Quality Control

In dealing with missing values and outliers in observed data, as well as default values in the numerical model, we adopt the difference method to fill gaps when there are less than or equal to 3 occurrences within 24 consecutive times. If there are more than 3 occurrences, the data for that day are not used.

2.5. LSTM Neural Network

2.5.1. Model Introduction

LSTM is a special type of RNN that is well suited for learning long time series information. Figure 2 illustrates the structural comparison between RNN and LSTM. It can be seen that, in an RNN structure, x_t represents the input information and h_t represents the output information. The traditional RNN network structure already has the capability to process time series data by transmitting processing information from previous moments to current moments and then on to subsequent moments. However, a limitation of RNN networks is that they can only receive information from adjacent sequence points, which may lead to issues such as gradient disappearing or gradient explosion when processing long sequence data.

To address this issue, LSTM replaces neural units in RNN with memory cells containing three “gates”—namely “input gates”, “output gates”, and “forgetting gates”. The key component of LSTM is its cell state represented by a horizontal line above each memory cell—similar to a conveyor belt running through the entire chain, allowing for downward flow of information. The “input gate”, “output gate”, and “forget gate” play crucial roles in selectively letting information through to protect and control the state of neural units by removing or adding output information from previous moments and input information from current moments into unit states.

The formulas involved in this structural diagram are as follows:

$$i_t=\sigma \left(W_i·\left[h_{t-1},x_t\right]+b_i\right)$$

(1)

$$f_t=\sigma \left(W_f·\left[h_{t-1},x_t\right]+b_f\right)$$

(2)

$$o_t=\sigma \left(W_o·\left[h_{t-1},x_t\right]+b_o\right)$$

(3)

$${\tilde{c}}_t=\tanh \left(W_c·\left[h_{t-1},x_t\right]+b_c\right)$$

(4)

$$c_t=f_t\times c_{t-1}+i_t\times {\tilde{c}}_t$$

(5)

$$h_t=o_t\times \tan \mathrm{h}\left(c_t\right)$$

(6)

In the formula, i_t, f_t, and o_t represent the “input gate”, “forgetting gate”, and “output gate” at time t, respectively; x_t represents the input information at time t; h_t−1 represents the output of the previous time; W and b are the corresponding weight coefficient matrix and offset top, respectively; σ and tanh denote the Sigmoid activation coefficient and hyperbolic tangent activation function, respectively; ${\tilde{c}}_t$ represents the temporary cell status; c_t represents the cell status update value at time t; and h_t is the output at time t.

After calculating the forgetting gate, input gate, and temporary cell status, the cell unit will update the cell status of the current moment. Finally, the output gate determines the output value ht of the current moment. More detailed information about LSTM can be found in reference [32].

2.5.2. Model Settings

After quality control, there are 1034 days of valid data from 1 August 2018 to 31 July 2021. The data are divided into two periods: 70% (725 days) for the training model and 30% (309 days) for the testing model. The objective of this paper is to solve the problem of short-term forecasting of SST for 3 days; therefore, we set the prediction length to 72 h in order to obtain time-by-time forecasting results for SST. To achieve better training results, the parameters of the LSTM model are set as follows through control experiments: numHiddenUnits are set to 200, MaxEpochs to 50, InitialLearnRate to 0.005 s, and LearnRateDropFactor to 0.2. Finally, in order to improve the stability and accuracy of the forecast, the ensemble forecast results of 10 members are used as the final SST forecast results (

Table 1. Parameter settings of LSTM model.

Parameters	Value
training set/%	70
test set/%	30
forecast time/hour	72
historical time/hour	72
numHiddenUnits	200
MaxEpochs	50
InitialLearnRate	0.005
LearnRateDropFactor	0.2
ensemble members	10

).

2.6. Test Indicators

Two indices, MAE and RMSE, were selected as indicators to measure the forecasting effect of the model using the following formula:

$$MAE={\displaystyle \frac{1}{\mathrm{m}}}{\displaystyle \sum _{i=1}^m\left|YPRE{D}_i-YTES{T}_i\right|}$$

(7)

$$RMSE=\sqrt{{\displaystyle \frac{1}{\mathrm{m}}}{\displaystyle \sum _{i=1}^m{(YPRE{D}_i-YTES{T}_i)}^2}}$$

(8)

where YPREDi represents the model’s prediction result for the ith sample, YTESTi represents the observed result for that sample, and m represents the number of samples used for testing.

3. Experimental Design

3.1. Experimental Scheme Setting

In this experiment, our focus is on predicting hourly changes in SST over a three-day period (represented as “Y”) at Xiaomaidao Ocean Station. We will be considering factors such as hourly observation values of SST over the past three days as well as future meteorological elements potentially related to SST changes at this station (represented as “X”). It includes air temperature at 2 m above sea level, relative humidity at 2 m above sea level, wind speed at 10 m above sea level, wind direction at 10 m above sea level, surface heat flux, latent heat flux, and air–sea temperature difference [33,34,35,36]. The basic information regarding these predictive factors is shown in

Table 2. Predictive factors for LSTM model.

ID	Variable	Abbreviation	Time Series	Data Source
①	Sea surface temperature	SST	(t − 71)~t 1	observation
②	Air temperature	AT	(t + 1)~(t + 72)	numerical model
③	Relative humidity	RH
④	Wind speed	WS
⑤	Wind direction	WD
⑥	Surface heat flux	HFx
⑦	Latent heat flux	LH
⑧	Air–sea temperature difference	AT-SST		observation and numerical mode

Note: ¹ t is the running time of the model, and t + 1 is the first time of the forecast.

.

In this experiment, SST is the forecast target, and the observation data before its start time are the basic information for training; therefore, SST is a mandatory factor for each group of experiments. When designing the experiments, EXP-1 is trained with SST as the only factor, meaning that only observed values of SST in the past are used to predict future SST changes. EXP-2 to EXP-7, respectively, added one meteorological factor such as air temperature, relative humidity, wind, etc., to SST observations. Based on tests on EXP-2 to EXP-7, the impacts of different meteorological factors on SST are discriminated. It should be noted that the wind vector consists of wind speed and direction, and the sea–air temperature difference refers to the difference between air temperature and the last observed SST. EXP-8 serves as a reserved experiment, which combines two meteorological factors that have had the greatest influence on SST selected from EXP-2 to EXP-7. In conclusion, a total of 8 groups of experiments are designed and each group undergoes 10 rounds of training, resulting in a total of 80 experiments (refer to

Table 3. Experimental scheme settings.

ID	Variable Combination	Training Times
EXP-1	SST	10
EXP-2	SST, AT	10
EXP-3	SST, RH	10
EXP-4	SST, WS, WD	10
EXP-5	SST, HFx	10
EXP-6	SST, LH	10
EXP-7	SST, (AT-SST)	10
EXP-8	optimal combination of EXP-2 to EXP-7	10

). If it turns out that minimal error occurs in EXP-8, then more diverse combinations of factors will be adopted for further experiments.

3.2. Experimental Process

Figure 3 illustrates the flowchart depicting the establishment process for the LSTM method prediction model. The specific steps include (1) reading quality controlled data from the file and standardizing it; (2) training the prediction factor (XTRAIN) based on LSTM in the training set to predict the target (YTRAIN), storing the trained neural network as “NET”; (3) calling “NET”, inputting the testing set’s prediction factor (XTEST), and calculating the prediction target (YPRED); (4) testing YPRED against YTEST in the testing set; (5) selecting the experiment with the smallest error across all experiments as the optimal prediction model (OPM). In the daily operational forecasting, the forecast value can be obtained by simply calling NET and inputting the value of the prediction factors.

4. Results and Tests

Based on Table 3′s experimental scheme settings, the model is trained and tested using an LSTM neural network. Figure 4 displays EXP-1′s daily and hourly test results. In Figure 4a,c, blue columns represent daily MAE and RMSE of ensemble member forecasts, while red columns represent ensemble forecasts of 10 members. It is evident that ensemble prediction errors are smaller than those of individual members, indicating that ensemble prediction based on LSTM models can enhance stability and accuracy compared to single models. In Figure 4b,d, black lines depict hourly MAE and RMSE of ensemble member forecasts, with red lines representing those of ensemble forecasts—further demonstrating improved stability and accuracy.

The same method was used to train EXP-2 through EXP-7 models; however, due to space constraints, only the best results after comparison are shown instead of listing each experiment’s test results like EXP-1.

Table 4. Test errors for EXP-1 to EXP-7.

	MAE			RMSE
ID	Day 1	Day 2	Day 3	Day 1	Day 2	Day 3
EXP-1	0.20	0.28	0.34	0.28	0.38	0.46
EXP-2	0.32	0.32	0.33	0.45	0.45	0.47
EXP-3	0.46	0.49	0.53	0.62	0.66	0.72
EXP-4	0.58	0.59	0.63	0.81	0.82	0.85
EXP-5	0.24	0.27	0.31	0.32	0.36	0.41
EXP-6	0.34	0.37	0.40	0.50	0.55	0.61
EXP-7	0.29	0.31	0.33	0.46	0.49	0.53

lists ensemble forecast test results for EXP-2 through EXP-7 as well as EXP-1. It is apparent that the overall effect is best for EXP-5 with errors on the second and third days smaller than those in EXP-1. The results of individual member forecasts versus ensemble forecasts of EXP-5 are depicted in Figure 5.

Based on the experimental scheme outlined above, we select the two experiments with the smallest experimental error from EXP-2 to EXP-7, combine their factors to form EXP-8, train the model, and compare its prediction effect.

Figure 6 illustrates the prediction errors of eight LSTM models (green columns) alongside those of a numerical model (yellow columns). The figure indicates that EXP-1 to EXP-8 yield much smaller prediction errors compared to those of the numerical model, demonstrating clear advantages of deep learning models in coastal ocean prediction. Specifically, for 1-day SST predictions, EXP-1 performs best followed by EXP-5; for 2-day SST predictions, EXP-5 excels followed by EXP-1; meanwhile, for 3-day SST predictions, EXP-5 demonstrates superior performance followed by EXP-2. Based on these findings, the OPM of SST for Xiaomaidao Ocean Station is constructed by combining the 1-day forecast from EXP-1 with the 2–3-day forecast from EXP-5. The forecast effect is depicted in Figure 7. The MAE values for 1–3 days using the OPM are 0.20 °C, 0.27 °C, and 0.31 °C, respectively, while the RMSE values are 0.28 °C, 0.36 °C, and 0.41 °C (Figure 7a,c). In terms of hourly forecast errors, the MAEs range between 0.10 °C and 0.40 °C for forecasts from the 1st hour to the 72nd hour, with RMSEs ranging between 0.20 °C and 0.50 °C (Figure 7b,d). On average, the OPM reduces forecast errors by as much as 78% compared to those of the numerical model.

5. Model Promotion

The method used to construct the OPM of SST at Xiaomaidao Ocean Station has been extended to 14 stations along the Bohai Sea and the Yellow Sea in order to improve forecasting accuracy across a wider area.

Another example, the Xiaoshidao Ocean Station, is used to demonstrate the forecasting performance of this method. Situated in the northeast of the Shandong Peninsula and facing the Yellow Sea to the north, the Xiaoshidao Ocean Station is approximately 220 km away from Xiaomaidao. As depicted in Figure 8, the forecasting errors of eight LSTM models are significantly smaller than that of the numerical model. Among them, Exp-5 has the smallest forecast error across 1–3 days, leading us to adopt the LSTM model trained by EXP-5 as the OPM for Xiaoshidao station. The MAEs for OPM range from 0.21 °C to 0.28 °C over a span of 1–3 days, while RMSEs range from 0.30 °C to 0.40 °C, decreasing by 76% compared with those produced by the numerical model.

Figure 9 illustrates the percentage improvement/reduction of the forecast effect at 14 stations in the Bohai Sea and the Yellow Sea. The results show that the coastal SST forecast error, when utilizing the LSTM method, is reduced by an average of 61% compared to the numerical model. Despite variations in the geographical location, surrounding environment, and different impact factors, it is evident that this method can enhance prediction accuracy to a certain degree when compared with the numerical model. Furthermore, it should be noted that the OPM running time obtained through the test is less than one minute, which significantly saves computing resources and obviously improves the forecast efficiency compared with the numerical model.

6. Summary and Discussion

In order to address large errors in predicting SST along coastlines using numerical models, this paper constructed SST prediction models for coastal stations in the Bohai Sea and the Yellow Sea based on LSTM—a type of deep learning network.

Firstly, Xiaomaidao Ocean Station was selected as an example to design an SST forecasting experiment. Factors related to SST changes—such as air temperature, wind vector, and heat flux—were extracted from the meteorological numerical model and combined with observed SST data to design different experimental schemes for LSTM model training. After testing forecast errors for each scheme, a combination yielding minimal error was selected as OPM. The 1–3-day MAEs of the OPM are 0.20 °C, 0.27 °C, and 0.31 °C, while the RMSEs are 0.28 °C, 0.36 °C, and 0.41 °C, respectively. In terms of hourly forecast errors, the MAEs range between 0.10 °C and 0.40 °C for forecasts from the 1st hour to the 72nd hour, with RMSEs ranging between 0.20 °C and 0.50 °C. When compared with the prediction results of the numerical model at the same time, it is found that the error of the OPM is reduced by an average of 78%.

The OPM construction method used for Xiaomaidao Ocean Station is extended to include 14 ocean stations along the Bohai Sea and the Yellow Sea. OPMs are constructed for each station and when compared with results from a numerical SST model for the same period, it is observed that on average, errors in predictions made by LSTM optimal models are 61% lower than those made by numerical models. This indicates that this method is universally applicable and can effectively improve coastal SST forecast accuracy. Similar studies have also been consulted. For instance, Zhang et al. developed an LSTM daily forecast model for SST in the equatorial Pacific (10° S–10° N, 120.0°–280° E) for the next 10 days, with an RMSE of 0.6 °C for the eastern equatorial Pacific and less than 0.3 °C for both central and western regions [37]. Han et al. utilized the LSTM model to predict daily SST at five buoy points in the East China Sea, with an MAE and RMSE of 0.25 °C and 0.28 °C for a one-day forecast, respectively [38]. The prediction errors of SST in these studies are similar to those found in this study, indicating that our constructed model is reasonable, reliable, and effective, especially considering the difficulty of predicting coastal SST compared to open sea SST. Furthermore, it is noted that the run time for all 14 stations using OPMs is less than one minute in total, which significantly saved computing resources and improved forecasting efficiency. Currently, this method has become a crucial reference for predicting SST in the Bohai Sea and the Yellow Sea. After an initial period of operation, it will be extended to a wider range of ocean stations in the future.

According to the OPM constructed at Xiaomaidao and Xiaoshidao ocean stations, as well as other stations, it is evident that the sea surface heat flux is the most significant factor influencing the change in SST. Following this, in terms of influence, are the sea–air temperature difference, latent heat flux, air temperature, relative humidity, and wind speed and direction. However, these factors are not orthogonal; that is, the factors affect each other. In our next step, we will consider performing the orthogonal decomposition of the influencing factors before screening them and then proceed to build a prediction model for the time series of each mode. Additionally, this study did not take into account oceanic factors such as tidal currents. Future research will consider these oceanic factors to enhance the accuracy of SST prediction. In terms of model building, we plan to integrate convolutional neural networks (CNNs) and LSTM to develop a hybrid model. The hybrid model could not only forecast the time series of SST but also incorporate linkage information between different sites.