DL-Mini Project (Ocean Wave Prediction) Karpagam K (2021188029)
DL-Mini Project (Ocean Wave Prediction) Karpagam K (2021188029)
DL-Mini Project (Ocean Wave Prediction) Karpagam K (2021188029)
i
f
Fig. 2. A series of plots illustrating fluctuations in ocean-
and
wave variables
The collected data were divided into two groups: one for
training and one for testing. The training group includes where 𝑥 <𝑡> is a vector from input matrix X at time t, 𝑔<𝑡>
most of the original data set and the test group consists is the information used to update cell state 𝑐 <𝑡> ,
what's left Next, the data was iterated through a divided superscript t is the current time index and t − 1 is the
into multiple NumPy arrays or sets. This process was previous time index. tanh is the hyperbolic tangent
repeated until all data was stored in the frame. As noted by function.
splitting the data sets into different ones frames uses
temporal information and improves the model
performance.The training dataset was used to build/train a Input weight matrices are 𝑊𝑖𝑔 , 𝑊𝑖𝑓 , 𝑊𝑖𝑖 , and 𝑊𝑖 ,with
deep learning model. Subsequently, a test data set was sizes N × B. B is the number of LSTM cells, 𝑊ℎ𝑔 , 𝑊ℎ𝑓 ,
delivered on a trained deep learning model to verify 𝑊ℎ𝑖 , and 𝑊ℎ𝑜 are the recurrent weight matrices of size N
performance compare with real-time data and assess × N.b(ex: 𝑏𝑜 )is the bias term in the activation function.
assessed training estimations by comparing training/
𝑓 <𝑡> , 𝑖 <𝑡> , 𝑜 <𝑡> are the forget, input, and output gates, validation losses, and training/ validation accuracies.
respectively. Losses were in terms of root mean squared error (RMSE)
and are equivalent to accuracies.
ℎ<𝑡> is the output from the LSTM layer.
where 𝑦𝑚^ is the prediction of interest m = 1, . . . , M at time Tables II and III lists the RMSEs (losses) for the training
t from output matrix 𝑦 ^ . and testing datasets, respectively
TABLE I TABLE II
RNN-LSTM DEEP LEARNING MODEL SUMMARY. TRAINING LOSS (RMSES).
Feature RMSE
Wvht 0.413
Layer (type) Shape Number of Dpd 12.55
Parameters Apd 1.123
lstm (LSTM) (None, 5, 32) 8065 Mwd 0.291
lstm 1 (LSTM) (None, 5, 16) 3136 Wtmp 0.413
dropout (None, 5, 16) 0
(Dropout)
lstm 2 (LSTM) (None, 10) 1080
dense (Dense) (None, 5) 55
TABLE III
TESTING LOSS (RMSE S).
E. HYPERPARAMETER OPTIMIZATION Feature RMSE
The Adam method was used to minimize the loss Wvht 0.394
function; dropout and batch normalization were Dpd 11.82
implemented to reduce overfitting and to hasten
Apd 1.089
convergence . In natural language processing models using
RNN, dropout rates range from 0.2 to 0.6 and 0.2 was used Mwd 0.307
in this study. We set the learning rate to 0.001, and a batch Wtmp 0.394
size to 200. Table I depicts the architecture of the deep
learning model.