3.3 Model Construction
According to the RNN network structure in Sect. 1.2 and the actual application
characteristics of maritime target trajectory prediction, in order to improve the prediction accuracy of the network, the appropriate timestep, the number of units in hidden
layer and the loss function need to be selected.
I. Timestep Selection
Timestep is an important parameter of time sequence prediction, which determines the
amount of accumulated historical information. The selection of timestep affects the
prediction accuracy of the network, small timestep will lead to insufficient learning of
trajectory characteristics, and large timestep will lead to gradient disappearance (explosion). In this paper, the target trajectory data at the 3–7 moment is selected as input,
the timestep is 3–7.
II. Selection of the Number of Units in Hidden Layer
The number of units is generally determined by the dimension of input data, which is
generally double dimension. Too few units leads to insufficient feature extraction, which
requires more training times and lower prediction accuracy. Too many units will lead to
overfitting. In this paper, the longitude, dimension and time interval of the maritime
target are selected as input, so the number of units in hidden layer is selected as 3–9.
III. Evaluation Index Selection
The evaluation index of the experiment determines the criterion for selecting the
optimal parameters and evaluating the prediction effect of the network. The evaluation
indexes for the test set are the sum of square error (SSE) and mean absolute error
(MAE), they reflect the dispersion and prediction accuracy.
The formulas of evaluation index are as follows:
MAE =
1
N
X N
i¼1
ðy
ðiÞ
À o
ðiÞ
Þ
ð 3:2Þ
SSE =
X N
i¼1
ðy
ðiÞ
À o
ðiÞ
Þ
2
ð3:3Þ
IV. Loss Function Selection
As an important part of network training, loss function is the learning criterion of
the whole network. Loss function reflects the inconsistency between predicted value
and real value. The smaller the value of the loss function is, the better performance of
the model will be. In this paper, Absolute loss (Formula 3.4) and Square loss (Formula 3.5) functions with good results are selected for comparison. Take RNN network
under different timestep as an example.
Maritime Target Trajectory Prediction Model Based on the RNN …
337
According to the RNN network structure in Sect. 1.2 and the actual application
characteristics of maritime target trajectory prediction, in order to improve the prediction accuracy of the network, the appropriate timestep, the number of units in hidden
layer and the loss function need to be selected.
I. Timestep Selection
Timestep is an important parameter of time sequence prediction, which determines the
amount of accumulated historical information. The selection of timestep affects the
prediction accuracy of the network, small timestep will lead to insufficient learning of
trajectory characteristics, and large timestep will lead to gradient disappearance (explosion). In this paper, the target trajectory data at the 3–7 moment is selected as input,
the timestep is 3–7.
II. Selection of the Number of Units in Hidden Layer
The number of units is generally determined by the dimension of input data, which is
generally double dimension. Too few units leads to insufficient feature extraction, which
requires more training times and lower prediction accuracy. Too many units will lead to
overfitting. In this paper, the longitude, dimension and time interval of the maritime
target are selected as input, so the number of units in hidden layer is selected as 3–9.
III. Evaluation Index Selection
The evaluation index of the experiment determines the criterion for selecting the
optimal parameters and evaluating the prediction effect of the network. The evaluation
indexes for the test set are the sum of square error (SSE) and mean absolute error
(MAE), they reflect the dispersion and prediction accuracy.
The formulas of evaluation index are as follows:
MAE =
1
N
X N
i¼1
ðy
ðiÞ
À o
ðiÞ
Þ
ð 3:2Þ
SSE =
X N
i¼1
ðy
ðiÞ
À o
ðiÞ
Þ
2
ð3:3Þ
IV. Loss Function Selection
As an important part of network training, loss function is the learning criterion of
the whole network. Loss function reflects the inconsistency between predicted value
and real value. The smaller the value of the loss function is, the better performance of
the model will be. In this paper, Absolute loss (Formula 3.4) and Square loss (Formula 3.5) functions with good results are selected for comparison. Take RNN network
under different timestep as an example.
Maritime Target Trajectory Prediction Model Based on the RNN …
337