vis

demo_prophet.py

@@ -22,6 +22,11 @@ def prophet_backend(train_data, forecast_horizon):
         yearly_seasonality=False, weekly_seasonality=True, daily_seasonality=True,
         holidays=HOLIDAY_DF)
 
+    # we can also play with setting daily_seasonality=False above, and then manually adding
+    # model.add_seasonality("daily", 1, fourier_order=10, prior_scale=100, mode="multiplicative")
+    # ...it didn't really work though. bumping the fourier_order helps, but makes the model slow.
+    # the rest didn't have much effect.
+
     model.fit(train_data)
 
     # Create a DataFrame with future timestamps for the evaluation period
@@ -34,17 +39,7 @@ def prophet_backend(train_data, forecast_horizon):
     for key in ('yhat', 'yhat_lower', 'yhat_upper'):
         forecast[key] = np.maximum(forecast[key], PREDICTION_LOWER_BOUND)
 
-    return forecast
-
-
-def sklearn_backend(train_data, forecast_horizon):
-    dc = train_data[['y']]
-    # inserting new column with yesterday's consumption values
-    for i in range(1, 4 * 24 + 1):
-        dc.loc[:, 'd%02d' % i] = dc.loc[:,'y'].shift(4 * 24 + i) # t-2days to t-1day
-        dc.loc[:, 'w%02d' % i] = dc.loc[:,'y'].shift(7 * 4 * 24 + i) # t-7days to t-8days
-    dc.info()
-    exit()
+    return forecast, model
 
 
 def prediction_task(backend, df, split_date, forecast_horizon):
@@ -53,12 +48,15 @@ def prediction_task(backend, df, split_date, forecast_horizon):
     eval_data = df[df['ds'] > split_date]
     eval_data = eval_data.head(forecast_horizon)
 
-    forecast = backend(train_data, forecast_horizon)
+    forecast, model = backend(train_data, forecast_horizon)
 
     mae = mean_absolute_error(eval_data['y'], forecast['yhat'])
 
     do_vis = False
     if do_vis:
+        future = model.make_future_dataframe(periods=forecast_horizon, freq='15T', include_history=True)
+        forecast = model.predict(future)
+
         plt.figure(figsize=(12, 6))
         plt.plot(eval_data['ds'], eval_data['y'], label='Actual', color='blue')
         plt.plot(forecast['ds'], forecast['yhat'], label='Predicted', color='red')
@@ -69,14 +67,14 @@ def prediction_task(backend, df, split_date, forecast_horizon):
         plt.legend()
         plt.grid(True)
         plt.show()
-        '''
+
         fig1 = model.plot(forecast)
         plt.plot(eval_data['ds'], eval_data['y'], c='r')
         plt.show()
-        '''
 
         fig2 = model.plot_components(forecast)
         plt.show()
+        exit()
 
     return mae, eval_data['y'].mean()
 
@@ -100,7 +98,8 @@ df['y'] = df['Consumption']
 
 
 # TODO 15 minutes timestep hardwired!
-forecast_horizon = 7 * 24 * 4
+forecast_horizon = 24 * 4
+print("forecast horizon", forecast_horizon // 4, "hours")
 
 
 start_date = '2021-06-01'
@@ -112,6 +111,7 @@ weekly_date_range = pd.date_range(start=start_date, end=end_date, freq='8d')
 maes = []
 mean_values = []
 for split_date in weekly_date_range:
+    # prophet_backend is the only backend currently
     mae, mean_value = prediction_task(prophet_backend, df, split_date, forecast_horizon)
     maes.append(mae)
     mean_values.append(mean_value)