|
|
|
|
|
|
|
import sklearn |
|
from tsai.basics import * |
|
|
|
|
|
import pandas as pd |
|
df=pd.read_csv('D:/project/aircompressordata.csv') |
|
df=df.drop(['data_id','ideal_specific_power_kw_100cfm','device_id'],axis=1) |
|
print(df.columns) |
|
df.head(5) |
|
|
|
|
|
|
|
|
|
|
|
datetime_col="timestamp" |
|
freq='1H' |
|
coloumns=df.columns[:7] |
|
method='ffill' |
|
value=0 |
|
|
|
preproc_pipe=sklearn.pipeline.Pipeline([ |
|
('shrinker',TSShrinkDataFrame()), |
|
('drop_duplicates',TSDropDuplicates(datetime_col=datetime_col)), |
|
('add_mts',TSAddMissingTimestamps(datetime_col=datetime_col,freq=freq)), |
|
('fill_missing',TSFillMissing(columns=coloumns,method=method,value=value)), |
|
], |
|
verbose=True) |
|
mkdir('data', exist_ok=True,parents=True) |
|
save_object(preproc_pipe,'data/preproc_pipe.pkl') |
|
preproc_pipe=load_object('data/preproc_pipe.pkl') |
|
|
|
df=preproc_pipe.fit_transform(df) |
|
|
|
|
|
|
|
|
|
|
|
df.head() |
|
|
|
|
|
|
|
|
|
|
|
fcst_history=200 |
|
fcst_horizon=72 |
|
valid_size=0.1 |
|
test_size=0.2 |
|
|
|
splits=get_forecasting_splits(df,fcst_history=fcst_history,fcst_horizon=fcst_horizon,datetime_col=datetime_col, |
|
valid_size=valid_size,test_size=test_size) |
|
|
|
splits |
|
|
|
|
|
|
|
|
|
|
|
coloumns=df.columns[1:] |
|
train_split=splits[0] |
|
|
|
exp_pipe=sklearn.pipeline.Pipeline([ |
|
('scaler',TSStandardScaler(columns=coloumns)), |
|
], |
|
verbose=True) |
|
|
|
save_object(exp_pipe,'data/exp_pipe.pkl') |
|
exp_pipe=load_object('data/exp_pipe.pkl') |
|
|
|
df_scaled=exp_pipe.fit_transform(df,scaler__idxs=train_split) |
|
|
|
df_scaled |
|
|
|
|
|
|
|
|
|
|
|
x_vars=df.columns[1:] |
|
y_vars=df.columns[1:] |
|
|
|
|
|
|
|
|
|
|
|
X,y=prepare_forecasting_data(df,fcst_history=fcst_history,fcst_horizon=fcst_horizon,x_vars=x_vars,y_vars=y_vars) |
|
X.shape , y.shape |
|
|
|
|
|
|
|
|
|
|
|
arch_config=dict( |
|
n_layers=3, |
|
n_heads=4, |
|
d_model=16, |
|
d_ff=128, |
|
attn_dropout=0.0, |
|
dropout=0.3, |
|
patch_len=24, |
|
stride=2, |
|
padding_patch=True, |
|
) |
|
|
|
|
|
|
|
|
|
|
|
learn=TSForecaster(X,y,splits=splits, |
|
batch_size=16,path="models", |
|
pipelines=[preproc_pipe,exp_pipe], |
|
arch="PatchTST", |
|
arch_config=arch_config, |
|
metrics=[mse,mae], |
|
cbs=ShowGraph()) |
|
|
|
|
|
|
|
|
|
|
|
learn.summary() |
|
|
|
|
|
|
|
|
|
|
|
lr_max=learn.lr_find().valley |
|
|
|
|
|
|
|
|
|
|
|
n_epochs=100 |
|
|
|
learn.fit_one_cycle(n_epoch=n_epochs,lr_max=lr_max) |
|
learn.export('PatchTST.pt') |
|
|
|
|
|
|
|
|
|
|
|
from tsai.inference import load_learner |
|
from sklearn.metrics import mean_squared_error, mean_absolute_error |
|
from sklearn.metrics import mean_absolute_percentage_error |
|
|
|
learn=load_learner('models/PatchTST.pt') |
|
y_test_preds, *_=learn.get_X_preds(X[splits[2]]) |
|
y_test_preds=to_np(y_test_preds) |
|
print(y_test_preds.shape) |
|
|
|
y_test=y[splits[2]] |
|
|
|
print(mean_squared_error(y_test.flatten(),y_test_preds.flatten())) |
|
print(mean_absolute_error(y_test.flatten(),y_test_preds.flatten())) |
|
print(mean_absolute_percentage_error(y_test.flatten(),y_test_preds.flatten())) |
|
|
|
|
|
|
|
|
|
|
|
X_test=X[splits[2]] |
|
plot_forecast(X_test,y_test,y_test_preds,sel_vars=True) |
|
|
|
|
|
|
|
|
|
|
|
fcst_date="2023-07-31 23:00:00" |
|
dates=pd.date_range(start=None,end=fcst_date,periods=fcst_history,freq=freq) |
|
dates |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
new_df=df[df[datetime_col].isin(dates)].reset_index(drop=True) |
|
new_df |
|
|
|
|
|
|
|
|
|
|
|
from tsai.inference import load_learner |
|
|
|
predict=load_learner('models/PatchTST.pt') |
|
new_df=predict.transform(new_df) |
|
|
|
new_df |
|
|
|
|
|
|
|
|
|
|
|
x_feat=new_df.columns[1:] |
|
new_x,__=prepare_forecasting_data(new_df,fcst_history=fcst_history,fcst_horizon=0,x_vars=x_vars,y_vars=y_vars) |
|
new_x.shape |
|
|
|
|
|
|
|
|
|
|
|
new_scaled_preds, *_ = learn.get_X_preds(new_x) |
|
|
|
new_scaled_preds=to_np(new_scaled_preds).swapaxes(1,2).reshape(-1,len(y_vars)) |
|
dates=pd.date_range(start=fcst_date, periods=fcst_horizon+1,freq='1H')[1:] |
|
preds_df=pd.DataFrame(dates,columns=[datetime_col]) |
|
preds_df.loc[:, y_vars]=new_scaled_preds |
|
preds_df=learn.inverse_transform(preds_df) |
|
|
|
preds_df |
|
|
|
|
|
|
|
|
|
|
|
from tsai.export import get_nb_name; nb_name=get_nb_name(locals()) |
|
from tsai.imports import create_scripts; create_scripts(nb_name) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
