python - 如何在一个简单的示例中使用 Python pytorch_forecasting 来预测时间序列？

Question

我想学习如何在一个简单的例子中使用pytorch_forecasting 。假设我们有一个只有 4 列的时间序列，即t,x(t),y(t),z(t). 这样的时间序列代表了粒子的轨迹。更具体地说：

import numpy  as np 
import pandas as pd 
import matplotlib.pyplot as plt
import os
nt=os.cpu_count(); #number of threads that my CPU offers

def plotS(xx,clr): #scatterplot of several sets of points with different colors
   xx=[( x if isinstance(x,np.ndarray) else 
         x.to_numpy() if isinstance(x,pd.DataFrame) else 
         x.numpy() if torch.is_tensor(x) 
         else 'Error!').astype('float32') for x in xx]
   fig=plt.figure(figsize=(15,10)); 
   ax=plt.axes(projection='3d'); 
   for i in range(len(xx)): 
      ax.scatter3D(xx[i][:,0],xx[i][:,1],xx[i][:,2],c=clr[i],s=4) #c=color
   x=np.concatenate(xx); 
   ax.set_box_aspect((np.ptp(x[:,0]),np.ptp(x[:,1]),np.ptp(x[:,2]))) #aspect ratio 1:1:1
   plt.tight_layout(); plt.show()

X=pd.DataFrame({'t':np.linspace(0,12*np.pi,10**4)}); #creating our dataset
#X['x(t)'], X['y(t)'], X['z(t)'] = np.cos(X['t'])*2*X['t'], np.sin(X['t'])*X['t'], 3*X['t']; #easy example
X['x(t)'], X['y(t)'], X['z(t)'] = np.sin(X['t']), np.sin(2*X['t']), X['t']/10 #hard example
X0,X1 = X.iloc[:int(0.7*len(X)),:], X.iloc[int(0.7*len(X)):,:] #split data into train and test part
plotS([X0.iloc[:,1:],X1.iloc[:,1:]],['b','g']) #visualize both parts

从前 70% 的点（蓝色部分），我希望预测最后 30%（绿色部分）。我对此的笨拙尝试pytorch_forecasting如下：

import torch, torchvision, pytorch_lightning, pytorch_forecasting
from pytorch_forecasting import Baseline, TemporalFusionTransformer, TimeSeriesDataSet

X['t']=np.arange(len(X)); #set the time to be consecutive integers
X['groups']=0;      #add a column, which marks all rows as coming from the same time-series
#X=X.drop([100,101,102,2000,2005,5123])
k,l=100,1;   tt=(X['t'].max()-X['t'].min());
y0= X.to_numpy()[X['t']<0.7*tt,1:4]; 
y1= X.to_numpy()[X['t']>0.7*tt,1:4]; 
X0= TimeSeriesDataSet(data=X[X['t']<0.7*tt],time_idx='t',target=['x(t)','y(t)','z(t)'], 
      group_ids=['groups'], max_encoder_length=k, max_prediction_length=l, allow_missing_timesteps=True,
      time_varying_known_reals=['t'],time_varying_unknown_reals=['x(t)','y(t)','z(t)']);   
X1=X[X['t']>0.7*tt]; #X1.iloc[:,1:]=0; 
X1= TimeSeriesDataSet.from_dataset(X0,X1,predict=False,stop_randomization=True); 
XX0=X0.to_dataloader(train=True, batch_size=64,num_workers=nt); 
XX1=X1.to_dataloader(train=False,batch_size=64,num_workers=nt); 
ann= TemporalFusionTransformer.from_dataset(X0,learning_rate=0.001,hidden_size=32,lstm_layers=4); 
trn=pytorch_lightning.Trainer(max_epochs=1,gpus=0);   
trn.fit(ann,XX0,XX0);  
y2=torch.cat(ann.predict(XX1),dim=1).numpy(); #cheating here: we use training values to predict
print(y1.shape,y2.shape); 
plotS([y0,y1,y2[:,:3]],['b','g','r']);

结果相当不错：

然而，计算是作弊的，因为ann.predict(XX1)它使用测试值来预测测试值。如何正确使用时间融合转换器，即仅从时间列预测测试值？

其次，我作弊是因为我将所有时间都转换为连续整数。如果我注释掉该行X['t']=np.arange(len(X))，我会得到错误。我如何预测时间序列不是整数。

第三，如果我将代码更改为k,l=100,10;，我会得到糟糕的结果：红色的预测变成了一条线。为什么？如何正确预测 l>1？

注意：我只是在学习 pytorch_forecasting，所以我很感激有关它的功能的一些提示。我确实阅读了文档，但没有简单的示例可以帮助我。

标签： pythonpytorchtime-seriesforecastingpytorch-lightning