"""
# Finding trend, seasonality, and remainder.
- Trend: Long term change.
- Seasonality: Repeated patterns over fixed interval such as daily.
- Remainder (irregular): Left from removing trend and seasonal components.

# 2 methods
1. Classic method: Can be multiplicative or additive where the sum/product of 
    trend, seasonality, and remainder recreate the original data
    - Logarithmic data is the same for both
    - Trend is estimated with a moving average
    - Seasonality estimated by averaging values for each period
2. Local regression: STL/MSTL uses LOESS (locally weighted scatterplot smothing)
    to be more flexible

Removing seasonality improves neural network forcasting performance
"""
import os
from statsmodels.tsa.seasonal import seasonal_decompose, STL, MSTL
import pandas as pd
import matplotlib.pyplot as plt

# Chart config
plt.rcParams['figure.figsize'] = (12, 6)

# Ensure assets directory
os.makedirs("assets", exist_ok=True)

# Load solar data
data = pd.read_csv(
    "../assets/datasets/time_series_solar.csv",
    parse_dates=["Datetime"],
    index_col="Datetime",
)
series = data["Incoming Solar"]

# Resample to daily
series_daily = series.resample("D").sum()

# Perform yearly seasonal decomposition with classical method
result = seasonal_decompose(x=series_daily, model='additive', period=365)
result.plot()
plt.savefig("assets/classical_decomposition.png")

# Perform yearly decomposition with Seasonal Trend decomposition using LOESS (STL)
result = STL(endog=series_daily, period=365).fit()
result.plot()
plt.savefig("assets/stl_decomposition.png")

# Perform yearly decomposition with mutliple STL (MSTL)
result = MSTL(endog=series_daily, periods=(7, 365)).fit()
result.plot()
plt.savefig("assets/mstl_decomposition.png")