Skip to main content

Table 1 Table showing the fixed parameters during evaluation, and the parameters, which are optimized by using grid search

From: A comprehensive study on battery electric modeling approaches based on machine learning

Method Fixed parameters Evaluated feature space Optimal
    constellation
MLR - - -
SVR RBF kernel w. γ=“scale” C=[1,...,80], C=30,
   ε =[0.008,...,0.032] ε=0.016
K-NN weights="uniform" p=[1,2,3], p=1,
  K-D tree k=[1,...,80] k=40
Decision Tree min_samples_leaf=5 Nleave=[100,...,30000] Nleave=10000
Random Forest Nleave=10000, Ntrees=[5,...,140] Ntrees=20
AdaBoost Nleave=10000, Ntrees=[5,...,140], Ntrees=40,
  loss=“linear” η=[0.5,...,2] η=2
GBR Nleave=31,η=0.1, Ntrees=[60,...,900] Ntrees=480
  loss=“least squares”   
FFNN activation=“ReLU” LFC=[1,2,3,4], LFC=3,
   NFC=[8,...,100] NFC=48
MergeFFNN activation=“ReLU” Ntop=[5,...,50], Ntop=35,
  NFC,a.m.=Ntop+Nbot Nbot=[10,...,40] Nbot=30
MergeLSTM NFC,a.m.=NLSTM+NFC NLSTM=[5,...,30], NLSTM=5,
   NFC=[15,...,40] NFC=25
LSTM - NLSTM=[5,...,35], NLSTM=15,
   NFC=[0,...,40] NFC=40
MergeCNN NFC,a.m.=Nfilter+NFC Nfilter=[5,...,25], Nfilter=15,
  filter_size,conv_Stride NFC=[10,...,40] NFC=40
CNN filter_size,conv_Stride Nfilter=[5,...,40], Nfilter=25,
   NFC=[0,...,50] NFC=35
  1. All neural networks use ADAM (Kingma and Ba 2017), a batch size of 256 and a learning rate η=0.0005. The parameter NFC,a.m. denotes the number of neurons in the dense layer after the merge layer