spinesTS

Time Series forecasting toolsets

• Install
• spinesTS Modules
• Tutorials
  • Getting started
  • Using nn module
    • StackingRNN
    • GAUNet
    • Time2VecNet
  • Using ml_model module
    • MultiStepRegressor
    • MultiOutputRegressor
    • WideGBRT
  • Using Data module

Install

pip install spinesTS

spinesTS Modules

• base: Model base class
• data: Built-in datasets and data wrapper classes
• feature_generator: Feature generation functions
• metrics: Model performance measurement function
• ml_model: Machine learning models
• nn: neural network models
• pipeline: Model fitting and prediction pipeline
• plotting: Visualization of model prediction results
• preprocessing: data preprocessing
• utils: Tool functions set
• layers: Neural network layer

Tutorials

Getting started

# simple demo to predict Electric data
from sklearn.preprocessing import StandardScaler
from lightgbm import LGBMRegressor
import matplotlib.pyplot as plt

from spinesTS.pipeline import Pipeline
from spinesTS.data import LoadElectricDataSets
from spinesTS.ml_model import MultiOutputRegressor
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    x_seq=df['value'], 
    y_seq=df['value'],  # The sequence of parameter y_seq is cut based on parameter x_seq
    # sliding window size, every 30 before days to predict after days
    window_size=30, 
    # predict after 30 days
    pred_steps=30, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# Assemble the model using Pipeline class
model = Pipeline([
    ('sc', StandardScaler()),
    ('model', MultiOutputRegressor(LGBMRegressor(random_state=2022)))
])
print("Model successfully initialization...")

# fitting model
model.fit(x_train, y_train, eval_set=(x_test, y_test), verbose=0)
print(f"r2_score is {model.score(x_test, y_test)}")

# plot the predicted results
fig = plot2d(y_test, model.predict(x_test), figsize=(20, 10), 
       eval_slices='[:30]', labels=['y_test', 'y_pred'])
plt.show()

[output]:
x_train shape is (270, 30), x_test shape is (68, 30),y_train shape is (270, 30),y_test shape is (68, 30)
Model successfully initialization...
r2_score is 0.8186046606725977

Using nn module

StackingRNN

import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d
from spinesTS.nn import StackingRNN
from spinesTS.metrics import r2_score, mean_absolute_error, mean_absolute_percentage_error


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    x_seq=df['value'], 
    y_seq=df['value'],
    # sliding window size, every 128 before days to predict after days
    window_size=128, 
    # predict after 24 days goods incoming
    pred_steps=24, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = StackingRNN(in_features=128, out_features=24, 
                    random_seed=42, loss_fn='mae', 
                    learning_rate=0.001, dropout=0.1, diff_n=1, 
                    stack_num=2, bidirectional=True, device='cpu')

model.fit(x_train, y_train, eval_set=(x_test[:-2], y_test[:-2]), batch_size=32,
             min_delta=0, patience=100, epochs=3000, verbose=False, lr_scheduler=None)
y_pred_cs = model.predict(x_test[-2:])
print(f"r2: {r2_score(y_test[-2:].T, y_pred_cs.T)}")
print(f"mae: {mean_absolute_error(y_test[-2:], y_pred_cs)}")
print(f"mape: {mean_absolute_percentage_error(y_test[-2:], y_pred_cs)}")
a = plot2d(y_test[-2:], y_pred_cs, eval_slices='[-1]', labels=['y_test', 'y_pred'], figsize=(20, 6))
plt.show()

GAUNet

import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d
from spinesTS.nn import GAUNet
from spinesTS.metrics import r2_score, mean_absolute_error, mean_absolute_percentage_error


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    x_seq=df['value'], 
    y_seq=df['value'],
    # sliding window size, every 128 before days to predict after days
    window_size=128, 
    # predict after 24 days 
    pred_steps=24, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = GAUNet(in_features=128, out_features=24, 
               random_seed=42, flip_features=False, 
               learning_rate=0.001, level=5, device='cpu')

model.fit(x_train, y_train, eval_set=(x_test[:-2], y_test[:-2]), batch_size=32,
             min_delta=0, patience=100, epochs=3000, verbose=False, lr_scheduler='ReduceLROnPlateau')
y_pred_cs = model.predict(x_test[-2:])
print(f"r2: {r2_score(y_test[-2:].T, y_pred_cs.T)}")
print(f"mae: {mean_absolute_error(y_test[-2:], y_pred_cs)}")
print(f"mape: {mean_absolute_percentage_error(y_test[-2:], y_pred_cs)}")
a = plot2d(y_test[-2:], y_pred_cs, eval_slices='[-1]', labels=['y_test', 'y_pred'], figsize=(20, 6))
plt.show()

Time2VecNet

import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d
from spinesTS.nn import Time2VecNet
from spinesTS.metrics import r2_score, mean_absolute_error, mean_absolute_percentage_error


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    x_seq=df['value'], 
    y_seq=df['value'],
    # sliding window size, every 128 before days to predict after days
    window_size=128, 
    # predict after 24 days 
    pred_steps=24, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = Time2VecNet(in_features=128, out_features=24, 
               random_seed=42, flip_features=False, 
               learning_rate=0.001, device='cpu')

model.fit(x_train, y_train, eval_set=(x_test[:-2], y_test[:-2]), batch_size=32,
             min_delta=0, patience=100, epochs=3000, verbose=False, lr_scheduler='CosineAnnealingLR')
y_pred_cs = model.predict(x_test[-2:])
print(f"r2: {r2_score(y_test[-2:].T, y_pred_cs.T)}")
print(f"mae: {mean_absolute_error(y_test[-2:], y_pred_cs)}")
print(f"mape: {mean_absolute_percentage_error(y_test[-2:], y_pred_cs)}")
a = plot2d(y_test[-2:], y_pred_cs, eval_slices='[-1]', labels=['y_test', 'y_pred'], figsize=(20, 6))
plt.show()

Using ml_model module

MultiStepRegressor

from lightgbm import LGBMRegressor
import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.ml_model import MultiStepRegressor
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    df['value'], 
    df['value'],
    # sliding window size, every 30 before days to predict after days
    window_size=30, 
    # predict after 30 days 
    pred_steps=30, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = MultiStepRegressor(LGBMRegressor(random_state=2022))
print("Model successfully initialization...")

# fitting model
model.fit(x_train, y_train, eval_set=(x_test, y_test), verbose=0)
print(f"r2_score is {model.score(x_test, y_test)}")

# plot the predicted results
fig = plot2d(y_test, model.predict(x_test), figsize=(20, 10), 
       eval_slices='[:30]', labels=['y_test', 'y_pred'])
plt.show()

MultiOutputRegressor

from lightgbm import LGBMRegressor
import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.ml_model import MultiOutputRegressor
from spinesTS.preprocessing import split_series
from spinesTS.plotting import plot2d


# load data
df = LoadElectricDataSets()

# split data
x_train, x_test, y_train, y_test = split_series(
    df['value'], 
    df['value'],
    # sliding window size, every 30 before days to predict after days
    window_size=30, 
    # predict after 30 days 
    pred_steps=30, 
    train_size=0.8
)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = MultiOutputRegressor(LGBMRegressor(random_state=2022))
print("Model successfully initialization...")

# fitting model
model.fit(x_train, y_train, eval_set=(x_test, y_test), verbose=0)
print(f"r2_score is {model.score(x_test, y_test)}")

# plot the predicted results
fig = plot2d(y_test, model.predict(x_test), figsize=(20, 10), 
       eval_slices='[:30]', labels=['y_test', 'y_pred'])
plt.show()

WideGBRT

from lightgbm import LGBMRegressor
import matplotlib.pyplot as plt

from spinesTS.data import LoadElectricDataSets
from spinesTS.ml_model import GBRTPreprocessing, WideGBRT
from spinesTS.plotting import plot2d


# load data
df = LoadElectricDataSets()

# split data and generate new features
gbrt_processor = GBRTPreprocessing(in_features=128, out_features=30, 
                                   target_col='value', train_size=0.8, date_col='date',
                                   differential_n=1  # The order of data differentiation.
                                   )
gbrt_processor.fit(df)

x_train, x_test, y_train, y_test = gbrt_processor.transform(df)

print(f"x_train shape is {x_train.shape}, "
      f"x_test shape is {x_test.shape}," 
      f"y_train shape is {y_train.shape},"
      f"y_test shape is {y_test.shape}")

# model initialization
model = WideGBRT(model=LGBMRegressor(random_state=2022))
print("Model successfully initialization...")

# fitting model
model.fit(x_train, y_train, eval_set=(x_test, y_test), verbose=0)
print(f"r2_score is {model.score(x_test, y_test)}")

# plot the predicted results
fig = plot2d(y_test, model.predict(x_test), figsize=(20, 10), 
       eval_slices='[:30]', labels=['y_test', 'y_pred'])
plt.show()

Using Data module

from spinesTS.data import *
series_data = BuiltInSeriesData(print_file_list=True)

+---+----------------------+----------------------------------------------+
|   | ds name              | columns                                      |
+---+----------------------+----------------------------------------------+
| 0 | ETTh1                | date, HUFL, HULL, MUFL, MULL, LUFL, LULL, OT |
| 1 | ETTh2                | date, HUFL, HULL, MUFL, MULL, LUFL, LULL, OT |
| 2 | ETTm1                | date, HUFL, HULL, MUFL, MULL, LUFL, LULL, OT |
| 3 | ETTm2                | date, HUFL, HULL, MUFL, MULL, LUFL, LULL, OT |
| 4 | Electric_Production  | date, value                                  |
| 5 | Messages_Sent        | date, ta, tb, tc                             |
| 6 | Messages_Sent_Hour   | date, hour, ta, tb, tc                       |
| 7 | Supermarket_Incoming | date, goods_cnt                              |
| 8 | Web_Sales            | date, type_a, type_b, sales_cnt              |
+---+----------------------+----------------------------------------------+

# select one dataset
df_a = series_data['ETTh1']  # series_data[0], it works, too
print(type(df_a))  # <class 'spinesTS.data._data_base.DataTS'>

# Because DataTS inherit from pandas DataFrame, it has all the functionality of pandas DataFrame
df_a.head() ,df_a.tail(), df_a.shape