scoring function improved

pvpower/estimator.py

@@ -28,7 +28,7 @@ class CoreVectorizer(Vectorizer):
 
     def vectorize(self, sample: WeatherForecast) -> List[float]:
         vectorized = [self._scale(sample.time_utc.month, 12),
-                      self._scale(12-abs(12-sample.time_utc.hour), 12),
+                      self._scale(sample.time_utc.hour, 24),
                       self._scale(sample.irradiance, 1000)]
         return vectorized
 
@@ -207,6 +207,7 @@ def num_samples_last_train(self) -> int:
     def retrain(self, train_data: TrainData):
         start = time.time()
         samples = train_data.samples
+        samples = [sample for sample in samples if sample.irradiance > 0]  # do not train the estimator for zero irradiance
         feature_vector_list = [self.__vectorizer.vectorize(sample) for sample in samples]
         label_list = [sample.power_watt for sample in samples]
         if len(set(label_list)) > 1:

pvpower/trainingcenter.py

@@ -15,9 +15,8 @@ def __init__(self, estimator: Estimator, train_data: TrainData):
         self.predictions = [estimator.predict(validation_record) for validation_record in self.validation_samples]
         self.estimator.retrain(train_data)  # retrain with all data
 
-    def __score(self, real, predicted):
-        distance = abs(real - predicted)
-        return distance
+    def __score(self, real, predicted)-> int:
+        return round(abs(real - predicted) / 10)*10
 
     @property
     def score(self) -> float:
@@ -28,8 +27,8 @@ def score(self) -> float:
             if real == 0 and predicted == 0:   # do not waste the total score by true zero predictions
                 continue
             scores.append(self.__score(real, predicted))
-        scores = self.__without_outliners(scores, 0.1)
-        return mean(scores)
+        scores_without_outliners = self.__without_outliners(scores, 0.1)
+        return mean(scores_without_outliners)
 
     @staticmethod
     def __without_outliners(scores: List[int], percent: float) -> List[int]: