Add model DSTAGNN

[hczs]

任务执行配置信息

数据集文件配置（DSTAGNNDataset.json），跑 PEMSD4 数据集时要调整 batch_size 参数为 32

{
  "batch_size": 64,
  "cache_dataset": true,
  "num_workers": 0,
  "train_rate": 0.6,
  "eval_rate": 0.2,
  "scaler": "standard",
  "load_external": false,
  "normal_external": false,
  "ext_scaler": "none",
  "input_window": 12,
  "output_window": 12,
  "sparsity": 0.01,
  "period": 288
}

模型参数配置（DSTAGNN.json）

{
  "max_epoch": 100,

  "learner": "adam",
  "learning_rate": 0.0001,
  "use_early_stop": false,

  "in_channels": 1,
  "nb_block": 4,
  "K": 3,
  "nb_chev_filter": 32,
  "nb_time_filter": 32,
  "d_model": 512,
  "d_k": 32,
  "n_heads": 3,
  "graph_use": "AG"
}

任务执行结果

模型评价指标对比如下，DSTAGNN 的指标是使用原论文代码本地复现的结果

• PEMSD4	PEMSD4	LibCity	DSTAGNN
  MAE	19.63	19.28
  RMSE	31.89	31.37
  MAPE	13.15%	12.72%

• PEMSD8	PEMSD8	LibCity	DSTAGNN
  MAE	16.11	15.81
  RMSE	25.15	25.03
  MAPE	10.44%	9.92%