A Novel Method for Regional Short-Term Forecasting of Water Level

: The water level forecasting system represented by the hydrodynamic model relies too much on the input data and the forecast value of the boundary, therefore introducing uncertainty in the prediction results. Tide tables ignore the effect of the residual water level, which is usually signiﬁcant. Therefore, to solve this problem, a water level forecasting method for the regional short-term (3 h) is proposed in this study. First, a simpliﬁed MIKE21 ﬂow model (FM) was established to construct the regional major astronomical tides after subdividing the model residuals into stationary constituents (surplus astronomical tides, simulation deviation) and nonstationary constituents (residual water level). Harmonic analysis (HA) and long short-term memory (LSTM) were adopted to forecast these model residuals, respectively. Finally, according to different spatial background information, the prediction for each composition was corrected by the inverse distance weighting (IDW) algorithm and its improved IDW interpolation algorithm based on signal energy and the spatial distance (IDWSE) from adjacent observation stations to nonmeasured locations. The developed method was applied to Narragansett Bay in Rhode Island. Compared with the assimilation model, the root-mean-square error (RMSE) of the proposed method decreased from 12.3 to 5.0 cm, and R 2 increased from 0.932 to 0.988. The possibility of adding meteorological features into the LSTM network was further explored as an extension of the prediction of the residual water level. The results show that the accuracy was limited to a moderate level, which is related to the difﬁculty presented by using only wind features to completely characterize the regional dynamic energy equilibrium process. if it is closer to the two observation stations along the coastline, and the value of interpolating point I 2 may be diminished if it is closer to the two observation stations far away from the coastline. scale, S M bathymetry hindcast astronomical tides original The results indicate that the variations in both the simulation deviation and surplus astronomical tides at any time can be forecasted by HA.


Introduction
As one of the most widely used approaches for tidal prediction, the tide tables based on harmonic analysis (HA) [1] can accurately predict astronomical tides. However, tide tables cannot predict the residual water level caused by wind, barometric pressure, and discharge, which are quite significant. With the development of numerical simulation technology, the effect of environmental factors on water level changes has become a concern. As a result, a series of hydrodynamic models have been proposed [2,3]. Relying on a hydrodynamic model and the Physical Oceanographic Real-Time System (PORTS ® ) [4], the United States built a 48 h tidal forecast system in ports, estuaries, great lakes, and coastal waters [5][6][7], such as the Gulf of Maine Operational Forecast System (GOMOFS). In this system, the Regional Ocean Modeling System (ROMs) [8] is used as the core prediction model, and meteorological and hydrological prediction products are also assimilated.
When evaluating the performance of GOMOFS, Peng et al. [9] found that the root-meansquare error (RMSE) between the prediction and the observation was larger than the RMSE between the observation and the astronomical tides prediction, indicating an unsatisfactory precision of prediction. The findings were the same in the Chesapeake Bay Operational Forecast System (CBOFS) [7].
Assimilation and correction of the Singapore regional model (SRM) [10] may be the work most similar to the research regional forecasting of water level. In these studies, prediction results from the model residuals are divided into two categories. One category is represented by chaos theory, in which the error is maintained at a medium level and does not change significantly [11]. For the second category, the error gradually diverges with lead time, such as in ensemble Kalman filters [12], neural networks [13], and other combined models [14]. Of note, as a product of numerical simulation, the forecast error always contains two tidal components [15]: The residuals of the simulated tides (simulation deviation) and the partial astronomical tides (surplus astronomical tides). The former is caused by the errors of bathymetry, coastline, parameters, open boundary, and so on, which will affect the total water depth and, in turn, affect the tide. The latter is caused by the number of tidal constituents at the open boundary, which are usually incomplete. For the prediction of nonstationary time series data, LSTM is considered to have excellent performance. Due to the dedicated "gate mechanism" [16], LSTM can remember variable lengths of time and has been successfully applied in the field of water level forecast, especially in marine disaster prevention (such as storm surges, floods) [17][18][19]. In fact, the short-time water level forecast under normal circumstances is also important for avoiding ship groundings, aiding in navigation and oil spill response. However, the related research is generally used for the whole water level at a single station, such as Yang et al. [20], and it may be unable to give the real forecasting skill of the LSTM for the residual water level.
Furthermore, the spatial distribution of the model residuals is also a crucial step for regional water level prediction. Clearly, a key ingredient in the successful spatial distribution is a realistic estimation of the background error distribution [21]. Wang et al. developed an approximate ordinary kriging method by hypothesizing that the spatial distribution of the model residuals is the same as that of the SRM output [11]. However, this assumption ignores the error caused by the simplified input and inaccurate parameters of the model itself, which change the spatial distribution of the model residuals. In the field of hydrology and meteorology, interpolation methods including inverse distance weighting (IDW) [22], kriging [23], and spline are commonly used. Considering the number of observation stations and the computational cost, it is more appropriate to use IDW to perform classic three-point interpolation (3PM) [24]. However, in some cases, the interpolation is amplified or diminished.
Therefore, we proposed a novel regional short-term (3 h) water level forecast method, which takes into account both the prediction accuracy of the residual water level at the observation station and the spatial distribution of water level components. A major astronomical tides model is constructed using the MIKE21 flow model (FM), and three constituents of model residuals are obtained by HA at the observation stations. The difficulty in model residual prediction lies in the residual water level. The LSTM is first used to forecast the residual water level under normal circumstances rather than extreme weather conditions, and its forecasting skill is comprehensively analyzed. As an extension of the prediction for residual water level, the possibility of adding meteorological features into the LSTM network for the prediction of residual water level is further explored. In addition, to solve the problem of weight anomalies when using the 3PM algorithm, the IDW interpolation method based on the signal energy and spatial distance (IDWSE) is proposed to improve the interpolation accuracy of the simulation deviation. Finally, to evaluate the performance of the proposed method, this method is applied to Narragansett Bay and the prediction results are compared with the assimilation model, which absorbs wind forcing and pressure.

Methods
The method used in this study is shown in Figure 1. The entire study can be divided into three parts: Construct a model of 8 major astronomical tides using the MIKE21 FM, forecast the model residuals at the observation stations, and distribute the forecasts at nonmeasured locations.
uate the performance of the proposed method, this method is applied to Narragansett Bay and the prediction results are compared with the assimilation model, which absorbs wind forcing and pressure.

Methods
The method used in this study is shown in Figure 1. The entire study can be divided into three parts: Construct a model of 8 major astronomical tides using the MIKE21 FM, forecast the model residuals at the observation stations, and distribute the forecasts at nonmeasured locations.

Numerical Model
MIKE 21 FM is a two-dimensional numerical model developed by the Danish Hydraulic Institute (DHI) for simulating tides, currents, waves, water quality, and other processes. The model has been successfully applied in rivers, lakes, estuaries, bays, and coastal areas [25][26][27][28]. The hydrodynamic module (HD) is based on the numerical solution of incompressible Reynolds averaged Navier-Stokes equations invoking the assumptions of Boussinesq and hydrostatic pressure and consists of continuity, momentum, temperature, salinity, and density equations [29]. In this study, MIKE21 FM is used to construct a model of major astronomical tides. Therefore, only the major astronomical tides are input at the open boundary.

Tidal Harmonic Analysis and Water Level Constituents
Decomposition and prediction of stationary constituents are realized by the HA, which determines the astronomical tide of a priori known frequencies (derived from astronomical and hydrodynamic theory): where S0 is the mean sea level (MSL) and R(t) is the residual water level. σ j , hj, gj, f, and

Numerical Model
MIKE 21 FM is a two-dimensional numerical model developed by the Danish Hydraulic Institute (DHI) for simulating tides, currents, waves, water quality, and other processes. The model has been successfully applied in rivers, lakes, estuaries, bays, and coastal areas [25][26][27][28]. The hydrodynamic module (HD) is based on the numerical solution of incompressible Reynolds averaged Navier-Stokes equations invoking the assumptions of Boussinesq and hydrostatic pressure and consists of continuity, momentum, temperature, salinity, and density equations [29]. In this study, MIKE21 FM is used to construct a model of major astronomical tides. Therefore, only the major astronomical tides are input at the open boundary.

Tidal Harmonic Analysis and Water Level Constituents
Decomposition and prediction of stationary constituents are realized by the HA, which determines the astronomical tide of a priori known frequencies (derived from astronomical and hydrodynamic theory): where S 0 is the mean sea level (MSL) and R(t) is the residual water level. σ j , h j , g j , f, and u are the angular velocity, amplitude, phase lag, amplitude, and phase correction factors corresponding to the jth tidal constituent, respectively. Of these, h j and g j are known to be the tidal harmonic constants. Equation (1) is linearized as: where h j = A j 2 + B j 2 / f j , g j = (V 0 + u) j + arctan(b j /a j ). The vectors a j and b j can be obtained by using least squares regression under the Rayleigh separation equation [30]. After obtaining the tidal harmonic constants (h j , g j ), the stationary constituents can be hindcasted or forecasted according to Equation (2).
The definition and properties of each water level constituent in this study are shown in Table 1.

Residual Water Level Forecast Based on the LSTM Network
Unlike feed-forward networks, recurrent neural networks (RNNs) utilize the internal memory to process arbitrary time sequences of inputs, for which there are both internal feedback and feed-forward connections between RNN cells [31]. LSTM is a modified version of RNN that can optionally add or remove information from the cell state (C t ) via the forget gate (f t ), input gate (i t ), and output gate (o t ) (Equations (3)-(8)). Thus, to some extent, LSTM solves the vanishing or exploding gradient problem of RNNs [6].
where W is the input weight matrix of the hidden element, U is the output weight matrix, b is the bias vector, '⊗' denotes the Hadamard product, and σ(·) is the activation function. In this study, the previous time step of length m is used to forecast the residual water level for the n-hour lead time. Next, the sample data set is generated by a single-step sliding window. In this way, the time sequences of the residual water level are reorganized and transformed into a supervised learning forecast problem. The LSTM structure is implemented in Keras with the TensorFlow backend [32], and the design concept of the network follows the principle of "simple to fine," that is, the hyperparameters are adjusted for the validation set so that the basic model with a simple structure can be updated to achieve an ideal outcome. The structure of the single-feature forecast model (SFFM) based on the LSTM network is shown in Figure 2.

Spatial Distribution Using IDW and IDWSE
IDW interpolation is commonly used in the geosciences and is usually applica situations with spatial distribution ambiguity. For IDW, it is assumed that weight de only on spatial distance rather than other physical processes. The expression is: where ˆ( ) c z t denotes the prediction at a nonmeasured location c; ( ) i z t denotes th diction at the ith observation station; q is the number of observation stations; di an are the Euclidean distance and weights between the nonmeasured location and ith o vation station, respectively; and p is the power exponent, which is 2 in the calculatio 3PM is an interpolation using three adjacent observation stations to estima value of nonmeasured locations. Thus, it is a fast and cost-effective method in tide-st insufficient water areas. The simulation deviation always increases from the open b ary to the bay in the semiclosed hydrodynamic model. Therefore, when the 3PM is for interpolation, the values of interpolation points are likely to be amplified or d ished, as shown in Figure 3.

Spatial Distribution Using IDW and IDWSE
IDW interpolation is commonly used in the geosciences and is usually applicable in situations with spatial distribution ambiguity. For IDW, it is assumed that weight depends only on spatial distance rather than other physical processes. The expression is: whereẑ c (t) denotes the prediction at a nonmeasured location c; z i (t) denotes the prediction at the ith observation station; q is the number of observation stations; d i and w i are the Euclidean distance and weights between the nonmeasured location and ith observation station, respectively; and p is the power exponent, which is 2 in the calculations. 3PM is an interpolation using three adjacent observation stations to estimate the value of nonmeasured locations. Thus, it is a fast and cost-effective method in tide-stationinsufficient water areas. The simulation deviation always increases from the open boundary to the bay in the semiclosed hydrodynamic model. Therefore, when the 3PM is used for interpolation, the values of interpolation points are likely to be amplified or diminished, as shown in Figure 3.
To address the abovementioned limitations, an interpolation method based on the signal energy and spatial distance (IDWSE) is proposed that is suitable for harmonic signals with linear spatial variation. The key idea of IDWSE is to correct the weak or strong weighting caused by 3PM through the signal energy ratio. First, the signal energy of the simulation deviation is calculated at each observation station (Equation (11)). Next, we identify station b with the minimum (or maximum) signal energy and calculate the signal energy multiple k i (at station b, k i = 1) of the residual observation stations relative to station b: where ε model,i (t) and E i denote the value and energy of the simulation deviation at the ith observation station, respectively. The values of the weights w 11 , w 12 , and w 13 can be expressed as: vation station, respectively; and p is the power exponent, which is 2 in the calculations. 3PM is an interpolation using three adjacent observation stations to estimate the value of nonmeasured locations. Thus, it is a fast and cost-effective method in tide-stationinsufficient water areas. The simulation deviation always increases from the open boundary to the bay in the semiclosed hydrodynamic model. Therefore, when the 3PM is used for interpolation, the values of interpolation points are likely to be amplified or diminished, as shown in Figure 3.  Finally, to prevent k i from being too large due to E i approaching zero, the following criteria are adopted to select the interpolation method:

Evaluation Index
To evaluate the accuracy of the forecast method, the following metrics are used in this study:

1.
Root-Mean-Square Error (RMSE) RMSE is one of the most widely used criteria for evaluating the accuracy of models. RMSE is very sensitive to a large forecast error, so it can measure the forecast performance of the high and low water levels well.

Mean Absolute Error (MAE)
MAE is a measure to evaluate the absolute deviation between the prediction and observation. The significance for the lower and higher water level forecasts is the same, and the deviation is not magnified by the square. 3.

R-Squared (R 2 )
R 2 describes the proportion of the total variance in the observation that can be explained by the forecast model. The closer R 2 is to 1, the better the regression results are.
Water 2021, 13, 820 7 of 17 In these equations, y t ,ŷ t , and y mean are the actual water level, predicted water level, and mean actual water level, respectively.

Experiment Area
The Narragansett Bay drainage basin covers an area of 4714 km 2 and is located on Rhode Island, USA. A total of 370 km 2 of this area is in Narragansett Bay (40 • 21 -41 • 51 N, 71 • 9 -71 • 30 W), which is the largest semiclosed estuary in the northeastern United States [33]. Fresh water flowing into the bay comes mainly from the Taunton, Blackstone, and Pawtuxet Rivers. However, compared with the tide, the magnitude of river discharge (combined yearly average of 105 m 3 s -1 ) is very small [34]. Seawater enters Narragansett Bay through three routes: The East Passage, West Passage, and Sakonnet River. The water depth in the East Passage is 16-48 m, while the West Passage is shallower (6-16 m) [35]. The diurnal range at Newport is 1.1 m and increases to 1.5 m at Providence.

Data Collection
The bathymetry and coastline data used by MIKE21 FM are derived from the products of Estuarine Bathymetry and Shoreline/Coastline Resources published by the National Oceanographic and Atmospheric Administration (NOAA), and the bathymetry datum is the mean lowest low water (MLLW). The tidal harmonic constant of the input water level at the open boundary is derived from the TPXO8_atlas tidal model published by the University of Oregon, USA [36]. The location of the five observation stations in the experimental area is shown in Figure 4. The hourly water level and meteorological observation from each station from 2013 to 2015 were selected for the experiment. To avoid introducing unnecessary errors in the datum transformation, the datum of the water level was vertically referenced to MSL in this paper.
The Narragansett Bay drainage basin covers an area of 4714 km 2 and is located on Rhode Island, USA. A total of 370 km 2 of this area is in Narragansett Bay ( 41 21 41 51 N 71 9 71 30 W --′ ′ ′ ′°°°°， ), which is the largest semiclosed estuary in the northeastern United States [33]. Fresh water flowing into the bay comes mainly from the Taunton, Blackstone, and Pawtuxet Rivers. However, compared with the tide, the magnitude of river discharge (combined yearly average of 105 m 3 s -1 ) is very small [34]. Seawater enters Narragansett Bay through three routes: The East Passage, West Passage, and Sakonnet River. The water depth in the East Passage is 16-48 m, while the West Passage is shallower (6-16 m) [35]. The diurnal range at Newport is 1.1 m and increases to 1.5 m at Providence.

Data Collection
The bathymetry and coastline data used by MIKE21 FM are derived from the products of Estuarine Bathymetry and Shoreline/Coastline Resources published by the National Oceanographic and Atmospheric Administration (NOAA), and the bathymetry datum is the mean lowest low water (MLLW). The tidal harmonic constant of the input water level at the open boundary is derived from the TPXO8_atlas tidal model published by the University of Oregon, USA [36]. The location of the five observation stations in the experimental area is shown in Figure 4. The hourly water level and meteorological observation from each station from 2013 to 2015 were selected for the experiment. To avoid introducing unnecessary errors in the datum transformation, the datum of the water level was vertically referenced to MSL in this paper.
Due to the limited number of observation stations, an experiment was designed to test the proposed method, in which Providence, Quonset Point (QP), and Fall River (FR) were used as observation stations, while Conimicut Light (CL) was regarded as the nonmeasured station to evaluate the experimental results. In addition, because of the absence of the water level at QP, six months (from April to September) of water level in 2015 were forecasted.  Due to the limited number of observation stations, an experiment was designed to test the proposed method, in which Providence, Quonset Point (QP), and Fall River (FR) were used as observation stations, while Conimicut Light (CL) was regarded as the nonmeasured station to evaluate the experimental results. In addition, because of the absence of the water level at QP, six months (from April to September) of water level in 2015 were forecasted.

Forecast Results for Stationary Constituents
A triangular irregular network (TIN) was used to construct the bathymetric mesh. To fit the coastline better, the resolution was increased along the coastline. The value of each mesh node was interpolated using high-resolution bathymetric data, as shown in Figure 5. The spatially varying bed roughness coefficient (Manning coefficient) and eddy viscosity coefficient (Smagorinsky coefficient) were typically adjusted as important initial parameters. In the absence of empirical parameters, these two constant coefficients were tested using an iterative approach using the major astronomical tides of Newport as the calibration target. Finally, the major astronomical tidal model of Narragansett Bay (MATNB) was constructed by using the parameters in Table 2, and it will be used in subsequent forecasting work.

Forecast Results for Stationary Constituents
A triangular irregular network (TIN) was used to construct the bathymetric mesh. To fit the coastline better, the resolution was increased along the coastline. The value of each mesh node was interpolated using high-resolution bathymetric data, as shown in Figure  5. The spatially varying bed roughness coefficient (Manning coefficient) and eddy viscosity coefficient (Smagorinsky coefficient) were typically adjusted as important initial parameters. In the absence of empirical parameters, these two constant coefficients were tested using an iterative approach using the major astronomical tides of Newport as the calibration target. Finally, the major astronomical tidal model of Narragansett Bay (MATNB) was constructed by using the parameters in Table 2, and it will be used in subsequent forecasting work.  Unlike the residual water level, the simulation deviation is still harmonic and gradually increases from the open boundary (Newport) to the inner bay, and reaches its maximum at the head of the bay, as shown in Table 3. The M2 constituent is the main component of the simulation deviation, which means that the amplitude of M2 is underestimated in MATNB. Table 4 presents the main components of the surplus astronomical tide. Two significant constituents are SA (long-period constituent) and M4 (shallow water constituent). Due to the large spatial scale, the SA is relatively close at each station. However, the amplitude of M4 is related to the bathymetry and tends to be larger in shallow water. The hindcast of the simulation deviation and the surplus astronomical tides account for 100% of the original signal variation. The results indicate that the variations in both the simulation deviation and surplus astronomical tides at any time can be forecasted by HA.  Unlike the residual water level, the simulation deviation is still harmonic and gradually increases from the open boundary (Newport) to the inner bay, and reaches its maximum at the head of the bay, as shown in Table 3. The M 2 constituent is the main component of the simulation deviation, which means that the amplitude of M 2 is underestimated in MATNB. Table 4 presents the main components of the surplus astronomical tide. Two significant constituents are S A (long-period constituent) and M 4 (shallow water constituent). Due to the large spatial scale, the S A is relatively close at each station. However, the amplitude of M 4 is related to the bathymetry and tends to be larger in shallow water. The hindcast of the simulation deviation and the surplus astronomical tides account for 100% of the original signal variation. The results indicate that the variations in both the simulation deviation and surplus astronomical tides at any time can be forecasted by HA.

Forecast Results for the Nonstationary Constituents
Before training the LSTM network, the sample was standardized by the z-score. Then, the two-year dataset (2013-2014) was divided into training and validation sets at a ratio of 7:3, and the next year's residual water level was used as the testing set to verify the prediction skill of the network. The hyperparameters of the network were adjusted for the validation set. To prevent overfitting, the training stopped when the loss stopped falling. The size of each dataset and hyperparameters are shown in Tables 5 and 6, respectively.  The RMSE and MAE statistics of the LSTM when the lead time (n) was equal to 3 are shown in Figure 6. The error was concentrated in the 10 cm range, with the lowest median prediction error in QP. Figure 7 further shows the performance of the LSTM network in the water level fluctuation period. The prediction at Providence does not fit the actual value well locally, showing slight underprediction of the high water level and slight overprediction of the low water level. By contrast, the performance at FR shows a moderate effect. However, the prediction performance is excellent at QP, which means that the high-frequency noise in shallow water interferes with the training and subsequently impacts the generalizability of the network. Regarding the overall forecast performance, the LSTM network still shows good stability and accuracy.
Water 2021, 13, x FOR PEER REVIEW 10 of 17 The RMSE and MAE statistics of the LSTM when the lead time (n) was equal to 3 are shown in Figure 6. The error was concentrated in the 10 cm range, with the lowest median prediction error in QP. Figure 7 further shows the performance of the LSTM network in the water level fluctuation period. The prediction at Providence does not fit the actual value well locally, showing slight underprediction of the high water level and slight overprediction of the low water level. By contrast, the performance at FR shows a moderate effect. However, the prediction performance is excellent at QP, which means that the highfrequency noise in shallow water interferes with the training and subsequently impacts the generalizability of the network. Regarding the overall forecast performance, the LSTM network still shows good stability and accuracy.   The RMSE and MAE statistics of the LSTM when the lead time (n) was equal to 3 are shown in Figure 6. The error was concentrated in the 10 cm range, with the lowest median prediction error in QP. Figure 7 further shows the performance of the LSTM network in the water level fluctuation period. The prediction at Providence does not fit the actual value well locally, showing slight underprediction of the high water level and slight overprediction of the low water level. By contrast, the performance at FR shows a moderate effect. However, the prediction performance is excellent at QP, which means that the highfrequency noise in shallow water interferes with the training and subsequently impacts the generalizability of the network. Regarding the overall forecast performance, the LSTM network still shows good stability and accuracy.

Spatial Distribution Results
As the main constituent in the simulation deviation, the amplitude of M 2 at Providence and FR is greater than that at QP, which causes weight anomalies when using 3PM interpolation. According to Table 7, the weights calculated using IDW and IDWSE have different biases. The performance of the two interpolation methods is shown in Figure 8. The curves show that the value at CL is amplified using IDW interpolation, while IDWSE interpolation solves this problem by increasing the weak weighting (QP). Furthermore, the interpolation accuracy is less sensitive to the interpolation method, which may be due to highly correlated components within the region. Despite this, IDWSE still improves the IDW interpolation accuracy slightly and proves its effectiveness.

Spatial Distribution Results
As the main constituent in the simulation deviation, the amplitude of M2 at Providence and FR is greater than that at QP, which causes weight anomalies when using 3PM interpolation. According to Table 7, the weights calculated using IDW and IDWSE have different biases. The performance of the two interpolation methods is shown in Figure 8. The curves show that the value at CL is amplified using IDW interpolation, while IDWSE interpolation solves this problem by increasing the weak weighting (QP). Furthermore, the interpolation accuracy is less sensitive to the interpolation method, which may be due to highly correlated components within the region. Despite this, IDWSE still improves the IDW interpolation accuracy slightly and proves its effectiveness.  In addition to the simulation deviation, the spatial backgrounds of other constituents were not clear, so IDW was still used for spatial correction. To better illustrate the sources of error in the water level forecasting, Figure 9 shows a time series of the astronomical tides and residual water level at different lead times at CL. The astronomical tides prediction was highly correlated with the actual values, indicating that the stationary constituent was well predicted and corrected (Figure 9a). The prediction of the residual water level was closely associated with the lead times (n). When n was equal to 1, the error was essentially distributed around zero (Figure 9b), and when n was equal to 3, the errors oscillated slightly (Figure 9c). The above results demonstrate that the accuracy of the water level forecasting at CL depends mainly on the accuracy of the residual water level forecasting at observation stations. In addition to the simulation deviation, the spatial backgrounds of other constituents were not clear, so IDW was still used for spatial correction. To better illustrate the sources of error in the water level forecasting, Figure 9 shows a time series of the astronomical tides and residual water level at different lead times at CL. The astronomical tides prediction was highly correlated with the actual values, indicating that the stationary constituent was well predicted and corrected (Figure 9a). The prediction of the residual water level was closely associated with the lead times (n). When n was equal to 1, the error was essentially distributed around zero (Figure 9b), and when n was equal to 3, the errors oscillated slightly (Figure 9c). The above results demonstrate that the accuracy of the water level forecasting at CL depends mainly on the accuracy of the residual water level forecasting at observation stations.

Comparison with Assimilation Model
To verify the superiority of the proposed method, we have developed an assimilation model based on MATNB for comparison. Just like the GOMOFS, this regional water level forecasting model assimilated the hourly variation in wind and barometric field. These data are obtained from ERA5 datasets [37] provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), which are available at a spatial resolution of 0.25° in longitude and latitude. Besides, eight major astronomical constituents (same as MATNB), as well as two shallow water overtides (M4 and MS4), were input at the open boundary.
The output of the assimilation model is shown in Figure 10a. Excluding the simulation start time, the prediction accuracy of the assimilation model for the high-water level was better than that of the low water level. In estuaries, the hazards of incorrectly estimated low water levels were more significant, which can cause the vessel groundings. Due to the limitations of the hydrodynamic model, this model cannot take into account all environmental factors. In other words, the forecasting error might be further magnified under extreme conditions. However, the proposed method captured the large oscillation process during the period of fluctuation, and the prediction errors were maintained at a low level during the stationary period despite the occasional outliers (Figure 10b). According to the statistical analysis (Table 8), the RMSE was reduced from 12.3 to 5.0 cm, the MAE was reduced from 9.7 to 3.8 cm, and the R 2 was increased to 0.988, indicating ideal forecast performance.

Comparison with Assimilation Model
To verify the superiority of the proposed method, we have developed an assimilation model based on MATNB for comparison. Just like the GOMOFS, this regional water level forecasting model assimilated the hourly variation in wind and barometric field. These data are obtained from ERA5 datasets [37] provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), which are available at a spatial resolution of 0.25 • in longitude and latitude. Besides, eight major astronomical constituents (same as MATNB), as well as two shallow water overtides (M 4 and MS 4 ), were input at the open boundary.
The output of the assimilation model is shown in Figure 10a. Excluding the simulation start time, the prediction accuracy of the assimilation model for the high-water level was better than that of the low water level. In estuaries, the hazards of incorrectly estimated low water levels were more significant, which can cause the vessel groundings. Due to the limitations of the hydrodynamic model, this model cannot take into account all environmental factors. In other words, the forecasting error might be further magnified under extreme conditions. However, the proposed method captured the large oscillation process during the period of fluctuation, and the prediction errors were maintained at a low level during the stationary period despite the occasional outliers (Figure 10b). According to the statistical analysis (Table 8), the RMSE was reduced from 12.3 to 5.0 cm, the MAE was reduced from 9.7 to 3.8 cm, and the R 2 was increased to 0.988, indicating ideal forecast performance.

Relationship between the Lead Time and Accuracy
As described previously, the residual water level is the primary component of prediction error. Thus, it is necessary to analyze the relationship between the lead time and accuracy in the LSTM network. Considering the moderate efficacy of the prediction performance at FR, we predicted the hourly residual water level at FR for 3, 6, 12, and 24 h lead time. The RMSE and R 2 of the prediction are illustrated in Figure 11. With increasing lead times, the overall performance of LSTM decreased slightly. Based on observations of the lth forecast time in each group, the RMSE entered the first significant growth interval with an increase of approximately 5 cm. Subsequently, the RMSE was basically unchanged when the stationary interval was 3 < l ≤ 6. The RMSE then entered a second significant growth interval at 6 < l ≤ 9, with an increase of approximately 3.5 cm. Thereafter, the RMSE increased slowly and remained near 13 cm, indicating that the prediction gradually converges. Using another precision index (R 2 ), its prediction was highly reliable (0.91) at the first forecast time. When l = 6, the predictions could still explain over 50% of the variation in the residual water level. Of note, R 2 became negative when l = 15, indicating a failure of the LSTM network.

Relationship between the Lead Time and Accuracy
As described previously, the residual water level is the primary component of prediction error. Thus, it is necessary to analyze the relationship between the lead time and accuracy in the LSTM network. Considering the moderate efficacy of the prediction performance at FR, we predicted the hourly residual water level at FR for 3, 6, 12, and 24 h lead time. The RMSE and R 2 of the prediction are illustrated in Figure 11. With increasing lead times, the overall performance of LSTM decreased slightly. Based on observations of the lth forecast time in each group, the RMSE entered the first significant growth interval with an increase of approximately 5 cm. Subsequently, the RMSE was basically unchanged when the stationary interval was 3 < l ≤ 6. The RMSE then entered a second significant growth interval at 6 < l ≤ 9, with an increase of approximately 3.5 cm. Thereafter, the RMSE increased slowly and remained near 13 cm, indicating that the prediction gradually converges. Using another precision index (R 2 ), its prediction was highly reliable (0.91) at the first forecast time. When l = 6, the predictions could still explain over 50% of the variation in the residual water level. Of note, R 2 became negative when l = 15, indicating a failure of the LSTM network. Water 2021, 13, x FOR PEER REVIEW 14 of 17

Comparison with the Multi-Feature Forecast Model Based on the LSTM Network
Wind and barometric pressure are generally considered significant factors that cause disturbances in the water level in the coastal ocean [38]. Therefore, we evaluated the possibility of using wind and barometric pressure as features (from NOAA at the observation stations) for residual water level prediction. To this end, the above features were entered into the LSTM along with the residual water level to build a multi-feature forecast model (MFFM). The dataset construction result is shown in Table 9.   Figure 12 shows the forecast performance of these two models with a lead time of 12 h at FR. According to the variation in the RMSE and R 2 curves, the interval l can be roughly divided into three parts: 0-3, 4-8, and 9-12 h. During the first interval [0, 3 h], meteorological features have a negative effect on the forecast. The RMSE of the MFFM at l = 1 reaches 6 cm, which is significantly larger than that of SFFM. During the second interval [4, 8 h], the forecast performance of these two models is almost equal. An interesting phenomenon is that the RMSE curve of MFFM does not continue to rise, but a "dent" occurs when l = 4 and 5. During the third interval [9, 12 h], the R 2 of the MFFM is greater than that of SFFM by approximately 0.1, indicating that the meteorological features improve the performance of the MFFM during this period.
It is well established that a component of water level variability is due to the inverted barometer effect [39]. To further analyze the error source of MFFM, we subtracted the barometric pressure effect from the residual water level using the following hydrostatic equation [40]: where η( ) t is the residual water level produced by barometric pressure (P(t)). ρ and g are the seawater density and acceleration due to gravity, respectively. Once the η( ) t is determined, the remaining components of residual water level (ψ(t)) are assumed to be produced by wind.

Comparison with the Multi-Feature Forecast Model Based on the LSTM Network
Wind and barometric pressure are generally considered significant factors that cause disturbances in the water level in the coastal ocean [38]. Therefore, we evaluated the possibility of using wind and barometric pressure as features (from NOAA at the observation stations) for residual water level prediction. To this end, the above features were entered into the LSTM along with the residual water level to build a multi-feature forecast model (MFFM). The dataset construction result is shown in Table 9. Table 9. Converting features to a dataset for the LSTM network.

Feat1(t)~Feat1
(t + n − 1) Residual water level Wind speed Wind direction Barometric pressure Residual water level 1 Feat1 (t − m)~Feat1 (t − 1) refers to feature 1 from time t − m to t − 1, and the same is true for the other features. Figure 12 shows the forecast performance of these two models with a lead time of 12 h at FR. According to the variation in the RMSE and R 2 curves, the interval l can be roughly divided into three parts: 0-3, 4-8, and 9-12 h. During the first interval [0, 3 h], meteorological features have a negative effect on the forecast. The RMSE of the MFFM at l = 1 reaches 6 cm, which is significantly larger than that of SFFM. During the second interval [4, 8 h], the forecast performance of these two models is almost equal. An interesting phenomenon is that the RMSE curve of MFFM does not continue to rise, but a "dent" occurs when l = 4 and 5. During the third interval [9, 12 h], the R 2 of the MFFM is greater than that of SFFM by approximately 0.1, indicating that the meteorological features improve the performance of the MFFM during this period.
It is well established that a component of water level variability is due to the inverted barometer effect [39]. To further analyze the error source of MFFM, we subtracted the barometric pressure effect from the residual water level using the following hydrostatic equation [40]: where η(t) is the residual water level produced by barometric pressure (P(t)). ρ and g are the seawater density and acceleration due to gravity, respectively. Once the η(t) is determined, the remaining components of residual water level (ψ(t)) are assumed to be produced by wind. We took the predicted value of l = 1 as a sequence and then used Equation (18) to obtain the 2 days' ψ(t) of MFFM at FR (Figure 13a). As a reference, we also provided the wind speed and direction, as shown in Figure 13b. The contrasts showed that MFFM could sometimes not accurately predict the direction and magnitude of the water level variability (within the dashed frame). In fact, due to the gravity and inertia, the water level is constantly changing to maintain a dynamic balance of the regional energy. Thus, it is difficult to fully characterize this process only by using wind features, and the accuracy of the MFFM is limited to a moderate level.

Conclusions
To improve the accuracy of the regional water level prediction, prediction and distribution of model residuals were carried out based on the simplified hydrodynamic model output (MATNB). Experimental results at Narragansett Bay showed that the SFFM can effectively predict the residual water level in the short-term (3 h). In terms of the spatial distribution, IDWSE improved the issue of amplitude the simulation deviation at CL using three-station IDW interpolation. Compared with the assimilation model, the RMSE of We took the predicted value of l = 1 as a sequence and then used Equation (18) to obtain the 2 days' ψ(t) of MFFM at FR (Figure 13a). As a reference, we also provided the wind speed and direction, as shown in Figure 13b. The contrasts showed that MFFM could sometimes not accurately predict the direction and magnitude of the water level variability (within the dashed frame). In fact, due to the gravity and inertia, the water level is constantly changing to maintain a dynamic balance of the regional energy. Thus, it is difficult to fully characterize this process only by using wind features, and the accuracy of the MFFM is limited to a moderate level.  We took the predicted value of l = 1 as a sequence and then used Equation (18) to obtain the 2 days' ψ(t) of MFFM at FR (Figure 13a). As a reference, we also provided the wind speed and direction, as shown in Figure 13b. The contrasts showed that MFFM could sometimes not accurately predict the direction and magnitude of the water level variability (within the dashed frame). In fact, due to the gravity and inertia, the water level is constantly changing to maintain a dynamic balance of the regional energy. Thus, it is difficult to fully characterize this process only by using wind features, and the accuracy of the MFFM is limited to a moderate level.

Conclusions
To improve the accuracy of the regional water level prediction, prediction and distribution of model residuals were carried out based on the simplified hydrodynamic model output (MATNB). Experimental results at Narragansett Bay showed that the SFFM can effectively predict the residual water level in the short-term (3 h). In terms of the spatial distribution, IDWSE improved the issue of amplitude the simulation deviation at CL using three-station IDW interpolation. Compared with the assimilation model, the RMSE of

Conclusions
To improve the accuracy of the regional water level prediction, prediction and distribution of model residuals were carried out based on the simplified hydrodynamic model output (MATNB). Experimental results at Narragansett Bay showed that the SFFM can effectively predict the residual water level in the short-term (3 h). In terms of the spatial distribution, IDWSE improved the issue of amplitude the simulation deviation at CL using three-station IDW interpolation. Compared with the assimilation model, the RMSE of this method decreased from 12.3 to 5.0 cm, and R 2 increased from 0.932 to 0.988. Therefore, this method could be a viable alternative for predicting the water level in the water area with few observation stations, and it does not require additional inputs of meteorological or hydrological prediction products.
Furthermore, we considered the influence of wind and barometric pressure on the LSTM network. The results showed that the accuracy of the MFFM was limited to a moderate level. This limitation arises because it is difficult to fully characterize the dynamic equilibrium process using only wind features. Therefore, we plan to design a reasonable feature quantification scheme to improve the prediction time and accuracy of the water level in future work.
Author Contributions: Z.T. conceptualized the study and wrote the paper with the guidance and supervision of D.S., X.G. and J.X., W.S. validated the results. Y.S. provided the software. All authors have read and agreed to the published version of the manuscript. Institutional Review Board Statement: Not applicable.