Research Article

DWNet: Dual-Window Deep Neural Network for Time Series Prediction

Table 2

MAPE and SMAPE performance comparison among different methods and datasets (best result is displayed in boldface).

ModelsSML 2010Human SportsEEGEnergy
MAPE (%)SMAPE (%)MAPE (%)SMAPE (%)MAPE (%)SMAPE (%)MAPE (%)SMAPE (%)

ARIMA (16)123.099362.009822.350718.4392159.634883.7905178.436577.4287
LSTM (16)78.909545.008217.443913.980380.342756.9819163.984965.8066
LSTM (128)83.102149.543917.903512.003387.560963.5271176.341569.5442
Encoder-Decoder (16)70.714243.076013.33269.561066.463538.5606170.086369.3110
Encoder-Decoder (128)78.998150.502215.798710.092379.344540.2327181.758376.1764
Input-Attn-RNN (16)61.112130.099711.78317.946241.885632.4032145.868867.5386
Input-Attn-RNN (128)68.908935.345912.00347.989741.462829.7608152.328764.7685
Temp-Attn-RNN (16)54.343532.870311.26277.211040.868329.0871140.683856.0774
Temp-Attn-RNN (128)57.806531.991111.49807.115345.870530.0085128.064459.3527
TCN (16)83.235049.579718.592011.0097675.9030131.4202258.170793.9882
TCN (128)85.447967.368914.614110.7326995.0083133.4580265.3023100.9782
LSTNet (16)50.595629.618613.09758.652446.920834.3753135.839668.8810
LSTNet (128)83.299946.306013.31929.369850.057341.5482140.202172.9974
DARNN (16)43.127528.955811.95688.068636.465826.6514123.055659.8798
DARNN (128)45.211033.161211.89517.417733.855027.1255139.068664.3326
DSTP (16)40.694624.526111.73597.134334.506322.7594138.895959.8884
DSTP (128)36.260024.804811.79287.340635.117924.0093142.874456.9903
DWNet31.576423.088810.38337.048331.328720.670682.111952.5880

The window size of baseline methods is set to 16 and 128, and the short window size and long window size of DWNet are set to 16 and 128, respectively.