Update predict.py. The prediction model is optimized to a time series model, whi…

…ch significantly improves the modeling fitness.

In the original method, only linear regression is used to perform simple trend extrapolation, which leads to insufficient prediction accuracy. This optimization adopts time series model, and uses the auto_arima method of pmdarima to automatically select appropriate model parameters (including p, d, q and seasonal parameters) according to historical data. It significantly improves the suitability of the model in time series modeling. In this way, the model can better capture the trend and periodicity of the data, and predict the future heat more reasonable and accurate.

 import numpy as np
 import datetime
-import matplotlib.pyplot as plt
+import pandas as pd
+from pmdarima import auto_arima
 
-
-def datetime_to_number(date: str):  # 格式化日期转换为 integer
+def datetime_to_number(date: str):
+    """Convert a date string 'YYYY-MM-DD' to a relative day number."""
     date_number = datetime.datetime.strptime(date, "%Y-%m-%d")
     base_number = datetime.datetime.strptime("2024-1-1", "%Y-%m-%d")
     return (date_number - base_number).days
 
+def predict_future_values(data, forecast_days=5):
+    """
+    Use auto_arima from pmdarima to fit a suitable ARIMA/SARIMA model for the time series,
+    then predict future values for the specified number of days.
+
+    Parameters:
+    data: dict, keys are date strings 'YYYY-MM-DD', values are integer counts
+    forecast_days: int, number of days to predict into the future
+
+    Returns:
+    predictions: dict, keys are future date strings 'YYYY-MM-DD', values are predicted integers (≥0)
+    """
+    if not data:
+        return {}
 
-def predict_future_values(data):
-    # 提取并排序日期
-    sorted_dates = sorted(data.keys(), key=lambda date: datetime.datetime.strptime(date, "%Y-%m-%d"))
-    sorted_data = {k: data[k] for k in sorted_dates}
+    # Sort data by date
+    sorted_dates = sorted(data.keys(), key=lambda d: datetime.datetime.strptime(d, "%Y-%m-%d"))
+    start_date = sorted_dates[0]
+    end_date = sorted_dates[-1]
     
-    # 将日期转换为整数并提取相应的值
-    xs = np.array([datetime_to_number(date) for date in sorted_data.keys()])
-    ys = np.array([data[date] for date in sorted_data.keys()])
+    # Create a full date range to ensure continuity in the time series
+    full_range = pd.date_range(start=start_date, end=end_date, freq='D')
+    ts = pd.Series(0, index=full_range, dtype=float)
+    for d in data:
+        ts[pd.to_datetime(d)] = data[d]
 
-    # 拟合线性回归模型
-    fit = np.polyfit(xs, ys, 1)
-    fn = np.poly1d(fit)
+    # Simple smoothing: optional step to reduce noise (moving average over 3 days)
+    # This is a mild smoothing to handle noisy data. You can comment this out if not needed.
+    ts_smoothed = ts.rolling(window=3, min_periods=1).mean()
     
-    # 获取最新日期，并生成未来三天的日期
-    latest_date = sorted_dates[-1]
-    latest_date_obj = datetime.datetime.strptime(latest_date, "%Y-%m-%d")
-    future_dates = [(latest_date_obj + datetime.timedelta(days=i)).strftime("%Y-%m-%d") for i in range(1, 6)]
+    # Fit the time series with auto_arima to find the best parameters
+    model = auto_arima(ts_smoothed, 
+                       start_p=1, start_q=1, 
+                       max_p=5, max_q=5, 
+                       seasonal=False,
+                       trace=False, error_action='ignore', suppress_warnings=True, stepwise=True)
     
-    # 预测未来日期的值
+    # Predict the future values
+    forecast = model.predict(n_periods=forecast_days)
+    # Construct future dates
+    last_date = pd.to_datetime(end_date)
+    future_dates = [last_date + datetime.timedelta(days=i) for i in range(1, forecast_days+1)]
+    
+    # Convert forecast results to dict with non-negative integers
     predictions = {}
-    for date in future_dates:
-        date_num = datetime_to_number(date)
-        if int(fn(date_num))<=0:
-            predictions[date] = 0
-        else:
-            predictions[date] = int(fn(date_num))
+    for d, v in zip(future_dates, forecast):
+        predictions[d.strftime("%Y-%m-%d")] = max(int(round(v)), 0)
 
     return predictions
 
-
 if __name__ == '__main__':
-    data = {'2024-06-15': 1, '2024-06-18': 1, '2024-06-22': 1, '2024-06-23': 1, '2024-07-01': 3, '2024-07-02': 4, '2024-07-03': 4, '2024-07-04': 14}
-    predictions = predict_future_values(data)
-    print(predictions)
-    # for date, value in predictions.items():
-    #     print(f'{date} PREDICTION: {value}')
+    data = {
+        '2024-06-15': 1, '2024-06-18': 1, '2024-06-22': 1,
+        '2024-06-23': 1, '2024-07-01': 3, '2024-07-02': 4,
+        '2024-07-03': 4, '2024-07-04': 14
+    }
+    preds = predict_future_values(data)
+    print(preds)