# xgboost修改版本 import os import pickle import pandas as pd import numpy as np from numpy.lib.stride_tricks import sliding_window_view import tkinter as tk import tkinter.font as tkfont from tkinter import ttk from datetime import timedelta from time import time import matplotlib.pyplot as plt from matplotlib.backends.backend_tkagg import FigureCanvasTkAgg, NavigationToolbar2Tk from xgboost import XGBRegressor from lunardate import LunarDate from sklearn.model_selection import train_test_split from sklearn.metrics import mean_squared_error, mean_absolute_error import matplotlib # 配置 matplotlib 中文显示 matplotlib.rcParams['font.sans-serif'] = ['SimHei', 'Microsoft YaHei', 'SimSun', 'Arial Unicode MS'] matplotlib.rcParams['axes.unicode_minus'] = False matplotlib.rcParams['font.family'] = 'sans-serif' # 全局缓存变量及特征名称 cached_model = None last_training_time = None feature_columns = None current_view = {'xlim': None, 'ylim': None} # 用于存储当前图表视图 # 数据加载与预处理函数 # ------------------------------- def load_data(upstream_file, downstream_file, river_level_file=None, flow_file=None): """ 加载所有相关数据并进行数据质量处理 """ try: # 读取上游和下游数据 upstream_df = pd.read_csv(upstream_file) downstream_df = pd.read_csv(downstream_file) except FileNotFoundError: print("文件未找到，请检查路径") return None # 确保列名一致 upstream_df.columns = ['DateTime', 'TagName', 'Value'] downstream_df.columns = ['DateTime', 'TagName', 'Value'] # 转换时间格式并设置为索引 upstream_df['DateTime'] = pd.to_datetime(upstream_df['DateTime']) downstream_df['DateTime'] = pd.to_datetime(downstream_df['DateTime']) # 设置DateTime为索引 upstream_df.set_index('DateTime', inplace=True) downstream_df.set_index('DateTime', inplace=True) # 数值处理 - 使用更稳健的转换方法 for df in [upstream_df, downstream_df]: df['Value'] = pd.to_numeric(df['Value'], errors='coerce') # 使用IQR方法检测异常值 Q1 = df['Value'].quantile(0.25) Q3 = df['Value'].quantile(0.75) IQR = Q3 - Q1 lower_bound = Q1 - 1.5 * IQR upper_bound = Q3 + 1.5 * IQR # 将异常值替换为边界值 df.loc[df['Value'] < lower_bound, 'Value'] = lower_bound df.loc[df['Value'] > upper_bound, 'Value'] = upper_bound # 过滤盐度小于5的数据 upstream_df = upstream_df[upstream_df['Value'] >= 5] downstream_df = downstream_df[downstream_df['Value'] >= 5] # 重命名Value列 upstream_df = upstream_df.rename(columns={'Value': 'upstream'})[['upstream']] downstream_df = downstream_df.rename(columns={'Value': 'downstream'})[['downstream']] # 合并数据 merged_df = pd.merge(upstream_df, downstream_df, left_index=True, right_index=True, how='inner') # 加载长江水位数据（如果提供） if river_level_file: try: river_level_df = pd.read_csv(river_level_file) print(f"成功读取水位数据文件: {river_level_file}") # 确保列名一致 if len(river_level_df.columns) >= 3: river_level_df.columns = ['DateTime', 'TagName', 'Value'] elif len(river_level_df.columns) == 2: river_level_df.columns = ['DateTime', 'Value'] river_level_df['TagName'] = 'water_level' # 数据处理 river_level_df['DateTime'] = pd.to_datetime(river_level_df['DateTime']) river_level_df.set_index('DateTime', inplace=True) river_level_df['Value'] = pd.to_numeric(river_level_df['Value'], errors='coerce') # 使用IQR方法处理异常值 Q1 = river_level_df['Value'].quantile(0.25) Q3 = river_level_df['Value'].quantile(0.75) IQR = Q3 - Q1 lower_bound = Q1 - 1.5 * IQR upper_bound = Q3 + 1.5 * IQR river_level_df.loc[river_level_df['Value'] < lower_bound, 'Value'] = lower_bound river_level_df.loc[river_level_df['Value'] > upper_bound, 'Value'] = upper_bound # 重命名并保留需要的列 river_level_df = river_level_df.rename(columns={'Value': 'water_level'})[['water_level']] # 合并到主数据框 merged_df = pd.merge(merged_df, river_level_df, left_index=True, right_index=True, how='left') # 对水位数据进行插值处理 merged_df['water_level'] = merged_df['water_level'].interpolate(method='time', limit=24) merged_df['water_level'] = merged_df['water_level'].fillna(method='ffill').fillna(method='bfill') # 创建平滑的水位数据 merged_df['water_level_smooth'] = merged_df['water_level'].rolling(window=24, min_periods=1, center=True).mean() # 添加水位趋势特征 merged_df['water_level_trend_1h'] = merged_df['water_level_smooth'].diff(1) merged_df['water_level_trend_24h'] = merged_df['water_level_smooth'].diff(24) print(f"水位数据加载成功，范围: {merged_df['water_level'].min()} - {merged_df['water_level'].max()}") except Exception as e: print(f"水位数据加载失败: {str(e)}") # 加载大通流量数据（如果提供） if flow_file: try: flow_df = pd.read_csv(flow_file) print(f"成功读取流量数据文件: {flow_file}") # 确保列名一致 if len(flow_df.columns) >= 3: flow_df.columns = ['DateTime', 'TagName', 'Value'] elif len(flow_df.columns) == 2: flow_df.columns = ['DateTime', 'Value'] flow_df['TagName'] = 'flow' # 数据处理 flow_df['DateTime'] = pd.to_datetime(flow_df['DateTime']) flow_df.set_index('DateTime', inplace=True) flow_df['Value'] = pd.to_numeric(flow_df['Value'], errors='coerce') # 使用IQR方法处理异常值 Q1 = flow_df['Value'].quantile(0.25) Q3 = flow_df['Value'].quantile(0.75) IQR = Q3 - Q1 lower_bound = Q1 - 1.5 * IQR upper_bound = Q3 + 1.5 * IQR flow_df.loc[flow_df['Value'] < lower_bound, 'Value'] = lower_bound flow_df.loc[flow_df['Value'] > upper_bound, 'Value'] = upper_bound # 重命名并保留需要的列 flow_df = flow_df.rename(columns={'Value': 'flow'})[['flow']] # 合并到主数据框 merged_df = pd.merge(merged_df, flow_df, left_index=True, right_index=True, how='left') # 对流量数据进行插值处理 merged_df['flow'] = merged_df['flow'].interpolate(method='time', limit=24) merged_df['flow'] = merged_df['flow'].fillna(method='ffill').fillna(method='bfill') # 创建平滑的流量数据 merged_df['flow_smooth'] = merged_df['flow'].rolling(window=24, min_periods=1, center=True).mean() # 添加流量趋势特征 merged_df['flow_trend_1h'] = merged_df['flow_smooth'].diff(1) merged_df['flow_trend_24h'] = merged_df['flow_smooth'].diff(24) # 添加流量统计特征 merged_df['mean_1d_flow'] = merged_df['flow_smooth'].rolling(window=24, min_periods=1).mean() merged_df['mean_3d_flow'] = merged_df['flow_smooth'].rolling(window=72, min_periods=1).mean() merged_df['std_1d_flow'] = merged_df['flow_smooth'].rolling(window=24, min_periods=1).std() # 添加流量变化特征 merged_df['flow_change_1h'] = merged_df['flow_smooth'].diff(1) merged_df['flow_change_24h'] = merged_df['flow_smooth'].diff(24) # # 添加流量与盐度比率（确保下游平滑数据已创建） # if 'downstream_smooth' in merged_df.columns: # merged_df['flow_sal_ratio'] = merged_df['flow_smooth'] / merged_df['downstream_smooth'] # else: # print("警告: 下游平滑数据未创建，跳过flow_sal_ratio计算") print(f"流量数据加载成功，范围: {merged_df['flow'].min()} - {merged_df['flow'].max()} m³/s") except Exception as e: print(f"流量数据加载失败: {str(e)}") # 对盐度数据进行插值和平滑处理 merged_df['upstream'] = merged_df['upstream'].interpolate(method='time', limit=24) merged_df['downstream'] = merged_df['downstream'].interpolate(method='time', limit=24) # 使用前向后向填充处理剩余的NaN值 merged_df['upstream'] = merged_df['upstream'].ffill().bfill() merged_df['downstream'] = merged_df['downstream'].ffill().bfill() # 创建平滑的盐度数据 merged_df['upstream_smooth'] = merged_df['upstream'].rolling(window=24, min_periods=1, center=True).mean() merged_df['downstream_smooth'] = merged_df['downstream'].rolling(window=24, min_periods=1, center=True).mean() # 添加趋势特征 merged_df['upstream_trend_1h'] = merged_df['upstream_smooth'].diff(1) merged_df['upstream_trend_24h'] = merged_df['upstream_smooth'].diff(24) merged_df['downstream_trend_1h'] = merged_df['downstream_smooth'].diff(1) merged_df['downstream_trend_24h'] = merged_df['downstream_smooth'].diff(24) # 填充NaN值 merged_df['upstream_trend_1h'] = merged_df['upstream_trend_1h'].fillna(0) merged_df['upstream_trend_24h'] = merged_df['upstream_trend_24h'].fillna(0) merged_df['downstream_trend_1h'] = merged_df['downstream_trend_1h'].fillna(0) merged_df['downstream_trend_24h'] = merged_df['downstream_trend_24h'].fillna(0) # 对低盐度部分使用更大的窗口进行平滑 low_sal_mask = merged_df['upstream'] < 50 if low_sal_mask.any(): merged_df.loc[low_sal_mask, 'upstream_smooth'] = merged_df.loc[low_sal_mask, 'upstream']\ .rolling(window=48, min_periods=1, center=True).mean() # 数据验证和统计 print("\n数据质量统计:") print(f"总数据量: {len(merged_df)}") print(f"上游盐度范围: {merged_df['upstream'].min():.2f} - {merged_df['upstream'].max():.2f}") print(f"下游盐度范围: {merged_df['downstream'].min():.2f} - {merged_df['downstream'].max():.2f}") if 'water_level' in merged_df.columns: print(f"水位范围: {merged_df['water_level'].min():.2f} - {merged_df['water_level'].max():.2f}") print(f"水位缺失比例: {merged_df['water_level'].isna().mean()*100:.2f}%") if 'flow' in merged_df.columns: print(f"流量范围: {merged_df['flow'].min():.2f} - {merged_df['flow'].max():.2f} m³/s") print(f"流量缺失比例: {merged_df['flow'].isna().mean()*100:.2f}%") # 重置索引，将DateTime作为列 merged_df = merged_df.reset_index() return merged_df # df = load_data('青龙港1.csv', '一取水.csv') # 测试 # df = load_data('青龙港1.csv', '一取水.csv') # df.to_csv('merged_data.csv', index=False) # print(f"Merged data saved to 'merged_data.csv' successfully") # # 绘制盐度随时间变化图 # plt.figure(figsize=(12, 6)) # plt.plot(df['DateTime'], df['upstream_smooth'], label='上游盐度', color='blue') # plt.plot(df['DateTime'], df['downstream_smooth'], label='下游盐度', color='red') # plt.xlabel('时间') # plt.ylabel('盐度') # plt.title('盐度随时间变化图') # plt.legend() # plt.grid(True) # plt.tight_layout() # plt.savefig('salinity_time_series.png', dpi=300) # plt.show() # ----------------------特征工程部分 # ------------------------------- # 添加农历（潮汐）特征 # ------------------------------- def add_lunar_features(df): lunar_day, lunar_phase_sin, lunar_phase_cos, is_high_tide = [], [], [], [] for dt in df['DateTime']: ld = LunarDate.fromSolarDate(dt.year, dt.month, dt.day) lunar_day.append(ld.day) lunar_phase_sin.append(np.sin(2 * np.pi * ld.day / 15)) lunar_phase_cos.append(np.cos(2 * np.pi * ld.day / 15)) is_high_tide.append(1 if (ld.day <= 5 or (ld.day >= 16 and ld.day <= 20)) else 0) df['lunar_day'] = lunar_day df['lunar_phase_sin'] = lunar_phase_sin df['lunar_phase_cos'] = lunar_phase_cos df['is_high_tide'] = is_high_tide return df # ------------------------------- # 生成延迟特征（向量化，利用 shift） # ------------------------------- def batch_create_delay_features(df, delay_hours): """ 为数据框中的特定列创建延迟特征 """ # 定义需要创建延迟特征的列 target_columns = ['upstream_smooth'] # target_columns = ['upstream_smooth', 'downstream_smooth'] # # 如果存在水位数据列，也为它创建延迟特征 暂时不使用 # if 'water_level_smooth' in df.columns: # target_columns.append('water_level_smooth') # elif 'water_level' in df.columns: # print("注意: 水位平滑列不存在，使用原始水位列创建延迟特征") # # 创建水位平滑列 # df['water_level_smooth'] = df['water_level'].rolling(window=24, min_periods=1, center=True).mean() # df['water_level_smooth'] = df['water_level_smooth'].fillna(df['water_level']) # target_columns.append('water_level_smooth') # 创建延迟特征 for column in target_columns: if column in df.columns: for delay in delay_hours: df[f'{column.split("_")[0]}_delay_{delay}h'] = df[column].shift(delay) else: print(f"警告: 列 {column} 不存在，跳过创建延迟特征") return df # 生成其他特征 def generate_features(df): """ 生成其他特征，包括历史数据、时间特征、统计特征和外部特征，并将这些特征添加到原始DataFrame中 """ try: # 创建平滑的盐度数据 df['upstream_smooth'] = df['upstream'].rolling(window=24, min_periods=1, center=True).mean() df['downstream_smooth'] = df['downstream'].rolling(window=24, min_periods=1, center=True).mean() # 时间特征 df['hour'] = df['DateTime'].dt.hour df['weekday'] = df['DateTime'].dt.dayofweek df['month'] = df['DateTime'].dt.month # 时间特征的sin和cos转换 df['hour_sin'] = np.sin(2 * np.pi * df['hour'] / 24) df['hour_cos'] = np.cos(2 * np.pi * df['hour'] / 24) df['weekday_sin'] = np.sin(2 * np.pi * df['weekday'] / 7) df['weekday_cos'] = np.cos(2 * np.pi * df['weekday'] / 7) df['month_sin'] = np.sin(2 * np.pi * df['month'] / 12) df['month_cos'] = np.cos(2 * np.pi * df['month'] / 12) # 统计特征 df['mean_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_up'] = df['upstream_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).std() df['mean_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_down'] = df['downstream_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).std() # 趋势特征 df['trend_1h_up'] = df['upstream_smooth'].diff(1) df['trend_3h_up'] = df['upstream_smooth'].diff(3) df['trend_6h_up'] = df['upstream_smooth'].diff(6) df['trend_12h_up'] = df['upstream_smooth'].diff(12) df['trend_24h_up'] = df['upstream_smooth'].diff(24) df['trend_1h_down'] = df['downstream_smooth'].diff(1) df['trend_3h_down'] = df['downstream_smooth'].diff(3) df['trend_6h_down'] = df['downstream_smooth'].diff(6) df['trend_12h_down'] = df['downstream_smooth'].diff(12) df['trend_24h_down'] = df['downstream_smooth'].diff(24) # 外部特征（水位和流量） if 'water_level_smooth' in df.columns: df['water_level_trend_1h'] = df['water_level_smooth'].diff(1) df['water_level_trend_24h'] = df['water_level_smooth'].diff(24) df['mean_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_water_level'] = df['water_level_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).std() if 'flow_smooth' in df.columns: df['flow_trend_1h'] = df['flow_smooth'].diff(1) df['flow_trend_24h'] = df['flow_smooth'].diff(24) df['mean_1d_flow'] = df['flow_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_flow'] = df['flow_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_flow'] = df['flow_smooth'].rolling(window=24, min_periods=1).std() return df except Exception as e: print(f"特征生成异常: {e}") return df # ------------------------------- # 向量化构造训练样本（优化特征工程） # ------------------------------- def create_features_vectorized(df, look_back=168, forecast_horizon=1): """ 向量化构造训练样本，使用过去7天的所有原始数据来预测未来1天的下游盐度均值 """ try: # 确保数据按时间排序 df = df.sort_values('DateTime') # 获取所有数值列（排除DateTime列） numeric_columns = df.select_dtypes(include=[np.number]).columns.tolist() if 'DateTime' in numeric_columns: numeric_columns.remove('DateTime') # 初始化特征和标签列表 features = [] # x输入 targets = [] # y输出 # 使用滑动窗口创建样本 for i in range(len(df) - look_back - forecast_horizon + 1): # 获取7天的特征窗口 window = df.iloc[i:i+look_back] # 提取特征 - 使用所有原始数据 window_features = [] for col in numeric_columns: # 获取列数据并处理NaN值 col_values = window[col].fillna(method='ffill').fillna(method='bfill').values window_features.extend(col_values) # 添加时间特征 current_date = window['DateTime'].iloc[-1] window_features.extend([ current_date.month, current_date.day, current_date.weekday() ]) # 获取目标值（未来1天的下游盐度均值） next_day = df.iloc[i+look_back:i+look_back+24] # 获取未来24小时的数据 # 处理目标值中的NaN target_values = next_day['downstream_smooth'].fillna(method='ffill').fillna(method='bfill').values target = np.mean(target_values) # 检查特征和目标值是否有效 if not np.any(np.isnan(window_features)) and not np.isnan(target) and not np.isinf(target): features.append(window_features) targets.append(target) if not features: print("警告: 未能生成任何有效特征") return np.array([]), np.array([]) # 转换为numpy数组 X = np.array(features) y = np.array(targets) print(f"成功生成特征矩阵，形状: {X.shape}") return X, y except Exception as e: print(f"特征创建异常: {e}") return np.array([]), np.array([]) def generate_prediction_features(df, current_date, look_back=168): """ 为预测生成特征，使用与create_features_vectorized相同的特征生成逻辑 """ try: # 确保数据按时间排序 df = df.sort_values('DateTime') # 获取所有数值列（排除DateTime列） numeric_columns = df.select_dtypes(include=[np.number]).columns.tolist() if 'DateTime' in numeric_columns: numeric_columns.remove('DateTime') # 找到当前日期在数据中的位置 current_idx = df[df['DateTime'] <= current_date].index[-1] # 获取过去168小时（7天）的数据窗口 if current_idx < look_back: print(f"数据不足，需要{look_back}小时的数据，但只有{current_idx+1}小时") return None window = df.iloc[current_idx-look_back+1:current_idx+1] # 提取特征 - 使用所有原始数据 features = [] for col in numeric_columns: # 直接使用原始数据作为特征 features.extend(window[col].values) # 添加时间特征 features.extend([ current_date.month, current_date.day, current_date.weekday() ]) return np.array(features) except Exception as e: print(f"预测特征生成异常: {e}") return None # ------------------------------- # 获取模型准确度指标 # ------------------------------- def get_model_metrics(): """获取保存在模型缓存中的准确度指标""" model_cache_file = 'salinity_model.pkl' if os.path.exists(model_cache_file): try: with open(model_cache_file, 'rb') as f: model_data = pickle.load(f) return { 'rmse': model_data.get('rmse', None), 'mae': model_data.get('mae', None) } except Exception as e: print(f"获取模型指标失败: {e}") return None # ------------------------------- # 模型训练与预测，展示验证准确度（RMSE, MAE） # ------------------------------- def train_and_predict(df, start_time, force_retrain=False): global cached_model, last_training_time model_cache_file = 'salinity_model.pkl' model_needs_training = True if os.path.exists(model_cache_file) and force_retrain: try: os.remove(model_cache_file) print("已删除旧模型缓存（强制重新训练）") except Exception as e: print("删除缓存异常:", e) train_df = df[df['DateTime'] < start_time].copy() # 创建测试特征，检查当前特征维度 test_X, test_y = create_features_vectorized(train_df, look_back=7, forecast_horizon=1) if test_X is None or test_y is None: print("特征生成失败") return None, None, None, None current_feature_dim = test_X.shape[1] if len(test_X) > 0 else 0 print(f"当前特征维度: {current_feature_dim}") cached_feature_dim = None if not force_retrain and cached_model is not None and last_training_time is not None: if last_training_time >= train_df['DateTime'].max(): try: cached_feature_dim = cached_model.n_features_in_ print(f"缓存模型特征维度: {cached_feature_dim}") if cached_feature_dim == current_feature_dim: model_needs_training = False print(f"使用缓存模型，训练时间: {last_training_time}") else: print(f"特征维度不匹配（缓存模型: {cached_feature_dim}，当前: {current_feature_dim}），需要重新训练") except Exception as e: print(f"检查模型特征维度失败: {e}") elif not force_retrain and os.path.exists(model_cache_file): try: with open(model_cache_file, 'rb') as f: model_data = pickle.load(f) cached_model = model_data['model'] last_training_time = model_data['training_time'] try: cached_feature_dim = cached_model.n_features_in_ print(f"文件缓存模型特征维度: {cached_feature_dim}") if cached_feature_dim == current_feature_dim: if last_training_time >= train_df['DateTime'].max(): model_needs_training = False print(f"从文件加载模型，训练时间: {last_training_time}") else: print(f"特征维度不匹配（文件模型: {cached_feature_dim}，当前: {current_feature_dim}），需要重新训练") except Exception as e: print(f"检查模型特征维度失败: {e}") except Exception as e: print("加载模型失败:", e) if model_needs_training: print("开始训练新模型...") if len(train_df) < 100: print("训练数据不足") return None, None, None, None start_train = time() X, y = create_features_vectorized(train_df, look_back=7, forecast_horizon=1) if X is None or y is None: print("特征生成失败") return None, None, None, None if len(X) == 0 or len(y) == 0: print("样本生成不足，训练终止") return None, None, None, None print(f"训练样本数量: {X.shape[0]}, 特征维度: {X.shape[1]}") X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.1, random_state=42) # 创建模型时设置 early_stopping_rounds model = XGBRegressor( n_estimators=200, learning_rate=0.1, max_depth=6, min_child_weight=2, subsample=0.8, colsample_bytree=0.8, gamma=0.1, reg_alpha=0.1, reg_lambda=1.0, n_jobs=-1, random_state=42, early_stopping_rounds=10 ) try: model.fit(X_train, y_train, eval_set=[(X_val, y_val)], eval_metric='rmse', verbose=False) # 在验证集上计算 RMSE 和 MAE y_val_pred = model.predict(X_val) rmse = np.sqrt(mean_squared_error(y_val, y_val_pred)) mae = mean_absolute_error(y_val, y_val_pred) print(f"验证集 RMSE: {rmse:.4f}, MAE: {mae:.4f}") # 特征重要性分析 feature_importance = model.feature_importances_ sorted_idx = np.argsort(feature_importance)[::-1] # 生成特征名称 feature_names = [] # 获取所有数值列 numeric_columns = train_df.select_dtypes(include=[np.number]).columns.tolist() if 'DateTime' in numeric_columns: numeric_columns.remove('DateTime') # 为每个数值列添加统计特征名称 for col in numeric_columns: feature_names.extend([ f'{col}_7d_mean_mean', f'{col}_7d_mean_std', f'{col}_7d_std_mean', f'{col}_7d_last_mean', f'{col}_7d_mean_change' ]) # 添加时间特征名称 feature_names.extend(['month', 'day', 'weekday']) # 确保特征名称数量与重要性数组长度匹配 if len(feature_names) != len(feature_importance): print(f"警告: 特征名称数量({len(feature_names)})与重要性数组长度({len(feature_importance)})不匹配") # 截取或填充特征名称以匹配重要性数组长度 feature_names = feature_names[:len(feature_importance)] # 打印前10个重要特征 print("\nTop 10 重要特征:") for i in range(min(10, len(sorted_idx))): print(f"{i+1}. {feature_names[sorted_idx[i]]}: {feature_importance[sorted_idx[i]]:.6f}") last_training_time = start_time cached_model = model with open(model_cache_file, 'wb') as f: pickle.dump({ 'model': model, 'training_time': last_training_time, 'feature_columns': feature_names, 'rmse': rmse, 'mae': mae, 'feature_dim': current_feature_dim }, f) print(f"模型训练完成，耗时: {time() - start_train:.2f}秒，特征维度: {current_feature_dim}") except Exception as e: print("模型训练异常:", e) return None, None, None, None else: model = cached_model # 预测部分 try: # 初始化存储预测结果的列表 future_dates = [start_time + timedelta(days=i) for i in range(5)] predictions = np.zeros(5) # 创建预测所需的特征矩阵 X_pred = [] for i in range(5): current_date = future_dates[i] features = generate_prediction_features(df, current_date, look_back=7) if features is None: print(f"生成预测特征失败: {current_date}") return None, None, None, None X_pred.append(features) # 批量预测 X_pred = np.array(X_pred) predictions = model.predict(X_pred) # 计算预测的置信区间 if model_needs_training: # 使用训练时的验证集误差 y_train_pred = model.predict(X_train) train_std = np.std(y_train - y_train_pred) else: # 使用模型缓存中的RMSE作为误差估计 try: with open(model_cache_file, 'rb') as f: model_data = pickle.load(f) train_std = model_data.get('rmse', 1.0) # 如果没有RMSE，使用默认值1.0 except: train_std = 1.0 # 如果无法获取RMSE，使用默认值1.0 prediction_intervals = np.array([ predictions - 1.96 * train_std, predictions + 1.96 * train_std ]) return future_dates, predictions, model, prediction_intervals except Exception as e: print("预测过程异常:", e) return None, None, None, None # ------------------------------- # GUI界面部分 # ------------------------------- def run_gui(): def configure_gui_fonts(): font_names = ['微软雅黑', 'Microsoft YaHei', 'SimSun', 'SimHei'] for font_name in font_names: try: default_font = tkfont.nametofont("TkDefaultFont") default_font.configure(family=font_name) text_font = tkfont.nametofont("TkTextFont") text_font.configure(family=font_name) fixed_font = tkfont.nametofont("TkFixedFont") fixed_font.configure(family=font_name) return True except Exception as e: continue return False def on_predict(): try: predict_start = time() status_label.config(text="预测中...") root.update() start_time_dt = pd.to_datetime(entry.get()) force_retrain = retrain_var.get() future_dates, predictions, model, prediction_intervals = train_and_predict(df, start_time_dt, force_retrain) if future_dates is None or predictions is None: status_label.config(text="预测失败") return # 获取并显示模型准确度指标 model_metrics = get_model_metrics() if model_metrics: metrics_text = f"模型准确度 - RMSE: {model_metrics['rmse']:.4f}, MAE: {model_metrics['mae']:.4f}" metrics_label.config(text=metrics_text) # 清除图形并重新绘制 ax.clear() # 创建双y轴图表 ax2 = None has_water_level = 'water_level' in df.columns and 'water_level_smooth' in df.columns if has_water_level: try: ax2 = ax.twinx() except Exception as e: print(f"创建双y轴失败: {e}") ax2 = None # 绘制历史数据（最近 120 天） history_end = min(start_time_dt, df['DateTime'].max()) history_start = history_end - timedelta(days=120) hist_data = df[(df['DateTime'] >= history_start) & (df['DateTime'] <= history_end)] # 确保数据不为空 if len(hist_data) == 0: status_label.config(text="错误: 所选时间范围内没有历史数据") return # 绘制基本数据 ax.plot(hist_data['DateTime'], hist_data['downstream_smooth'], label='一取水(下游)盐度', color='blue', linewidth=1.5) ax.plot(hist_data['DateTime'], hist_data['upstream_smooth'], label='青龙港(上游)盐度', color='purple', linewidth=1.5, alpha=0.7) # 绘制水位数据（如果有） if ax2 is not None and has_water_level: try: # 检查水位数据是否有足够的非NaN值 valid_water_level = hist_data['water_level_smooth'].dropna() if len(valid_water_level) > 10: # 至少有10个有效值 ax2.plot(hist_data['DateTime'], hist_data['water_level_smooth'], label='长江水位', color='green', linewidth=1.5, linestyle='--') ax2.set_ylabel('水位 (m)', color='green') ax2.tick_params(axis='y', labelcolor='green') else: print("水位数据有效值不足，跳过水位图") except Exception as e: print(f"绘制水位数据时出错: {e}") # 绘制预测数据 if len(future_dates) > 0 and len(predictions) > 0: ax.plot(future_dates, predictions, marker='o', linestyle='--', label='递归预测盐度', color='red', linewidth=2) # 添加预测的置信区间 if prediction_intervals is not None: ax.fill_between(future_dates, prediction_intervals[0], prediction_intervals[1], color='red', alpha=0.2, label='95% 置信区间') # 绘制实际数据(如果有) actual_data = df[(df['DateTime'] >= start_time_dt) & (df['DateTime'] <= future_dates[-1])] actual_values = None if not actual_data.empty: actual_values = [] # 获取与预测日期最接近的实际数据 for pred_date in future_dates: closest_idx = np.argmin(np.abs(actual_data['DateTime'] - pred_date)) actual_values.append(actual_data['downstream_smooth'].iloc[closest_idx]) # 绘制实际盐度曲线 ax.plot(future_dates, actual_values, marker='s', linestyle='-', label='实际盐度', color='orange', linewidth=2) # 设置图表标题和标签 ax.set_xlabel('日期') ax.set_ylabel('盐度') ax.set_title(f"从 {start_time_dt.strftime('%Y-%m-%d %H:%M:%S')} 开始的递归单步盐度预测") # 设置图例并应用紧凑布局 if ax2 is not None: try: lines1, labels1 = ax.get_legend_handles_labels() lines2, labels2 = ax2.get_legend_handles_labels() if lines2: # 确保水位数据已绘制 ax.legend(lines1 + lines2, labels1 + labels2, loc='best') else: ax.legend(loc='best') except Exception as e: print(f"创建图例时出错: {e}") ax.legend(loc='best') else: ax.legend(loc='best') fig.tight_layout() # 强制重绘 plt.close(fig) fig.canvas.draw() fig.canvas.flush_events() plt.draw() # 更新预测结果文本 predict_time = time() - predict_start status_label.config(text=f"递归预测完成 (耗时: {predict_time:.2f}秒)") # 显示预测结果 result_text = "递归单步预测结果:\n\n" # 如果有实际值，计算差值和百分比误差 if actual_values is not None: result_text += "日期 预测值 实际值 差值\n" result_text += "--------------------------------------\n" for i, (date, pred, actual) in enumerate(zip(future_dates, predictions, actual_values)): diff = pred - actual result_text += f"{date.strftime('%Y-%m-%d')} {pred:6.2f} {actual:6.2f} {diff:6.2f}\n" else: result_text += "日期 预测值\n" result_text += "-------------------\n" for i, (date, pred) in enumerate(zip(future_dates, predictions)): result_text += f"{date.strftime('%Y-%m-%d')} {pred:6.2f}\n" result_text += "\n无实际值进行对比" update_result_text(result_text) except Exception as e: status_label.config(text=f"错误: {str(e)}") import traceback traceback.print_exc() def on_scroll(event): xlim = ax.get_xlim() ylim = ax.get_ylim() zoom_factor = 1.1 x_data = event.xdata if event.xdata is not None else (xlim[0]+xlim[1])/2 y_data = event.ydata if event.ydata is not None else (ylim[0]+ylim[1])/2 x_rel = (x_data - xlim[0]) / (xlim[1] - xlim[0]) y_rel = (y_data - ylim[0]) / (ylim[1] - ylim[0]) if event.step > 0: new_width = (xlim[1]-xlim[0]) / zoom_factor new_height = (ylim[1]-ylim[0]) / zoom_factor x0 = x_data - x_rel * new_width y0 = y_data - y_rel * new_height ax.set_xlim([x0, x0+new_width]) ax.set_ylim([y0, y0+new_height]) else: new_width = (xlim[1]-xlim[0]) * zoom_factor new_height = (ylim[1]-ylim[0]) * zoom_factor x0 = x_data - x_rel * new_width y0 = y_data - y_rel * new_height ax.set_xlim([x0, x0+new_width]) ax.set_ylim([y0, y0+new_height]) canvas.draw_idle() def update_cursor(event): if event.inaxes == ax: canvas.get_tk_widget().config(cursor="fleur") else: canvas.get_tk_widget().config(cursor="") def reset_view(): display_history() status_label.config(text="图表视图已重置") root = tk.Tk() root.title("青龙港-陈行盐度预测系统") try: configure_gui_fonts() except Exception as e: print("字体配置异常:", e) # 恢复输入框和控制按钮 input_frame = ttk.Frame(root, padding="10") input_frame.pack(fill=tk.X) ttk.Label(input_frame, text="输入开始时间 (YYYY-MM-DD HH:MM:SS)").pack(side=tk.LEFT) entry = ttk.Entry(input_frame, width=25) entry.pack(side=tk.LEFT, padx=5) predict_button = ttk.Button(input_frame, text="预测", command=on_predict) predict_button.pack(side=tk.LEFT) status_label = ttk.Label(input_frame, text="提示: 第一次运行请勾选'强制重新训练模型'") status_label.pack(side=tk.LEFT, padx=10) control_frame = ttk.Frame(root, padding="5") control_frame.pack(fill=tk.X) retrain_var = tk.BooleanVar(value=False) ttk.Checkbutton(control_frame, text="强制重新训练模型", variable=retrain_var).pack(side=tk.LEFT) # 更新图例说明，加入水位数据信息 if 'water_level' in df.columns: legend_label = ttk.Label(control_frame, text="图例: 紫色=青龙港上游数据, 蓝色=一取水下游数据, 红色=预测值, 绿色=长江水位") else: legend_label = ttk.Label(control_frame, text="图例: 紫色=青龙港上游数据, 蓝色=一取水下游数据, 红色=预测值, 橙色=实际值") legend_label.pack(side=tk.LEFT, padx=10) reset_button = ttk.Button(control_frame, text="重置视图", command=reset_view) reset_button.pack(side=tk.LEFT, padx=5) # 添加显示模型准确度的标签 metrics_frame = ttk.Frame(root, padding="5") metrics_frame.pack(fill=tk.X) model_metrics = get_model_metrics() metrics_text = "模型准确度: 未知" if not model_metrics else f"模型准确度 - RMSE: {model_metrics['rmse']:.4f}, MAE: {model_metrics['mae']:.4f}" metrics_label = ttk.Label(metrics_frame, text=metrics_text) metrics_label.pack(side=tk.LEFT, padx=10) # 结果显示区域 result_frame = ttk.Frame(root, padding="10") result_frame.pack(fill=tk.BOTH, expand=True) # 左侧放置图表 plot_frame = ttk.Frame(result_frame, width=800, height=600) plot_frame.pack(side=tk.LEFT, fill=tk.BOTH, expand=True) plot_frame.pack_propagate(False) # 不允许框架根据内容调整大小 # 右侧放置文本结果 text_frame = ttk.Frame(result_frame) text_frame.pack(side=tk.RIGHT, fill=tk.Y) # 使用等宽字体显示结果 result_font = tkfont.Font(family="Courier New", size=10, weight="normal") # 添加文本框和滚动条 result_text = tk.Text(text_frame, width=50, height=25, font=result_font, wrap=tk.NONE) result_text.pack(side=tk.LEFT, fill=tk.BOTH) result_scroll = ttk.Scrollbar(text_frame, orient="vertical", command=result_text.yview) result_scroll.pack(side=tk.RIGHT, fill=tk.Y) result_text.configure(yscrollcommand=result_scroll.set) result_text.configure(state=tk.DISABLED) # 初始设为只读 # 更新结果文本的函数 def update_result_text(text): result_text.configure(state=tk.NORMAL) result_text.delete(1.0, tk.END) result_text.insert(tk.END, text) result_text.configure(state=tk.DISABLED) # 创建更高DPI的图形以获得更好的显示质量 fig, ax = plt.subplots(figsize=(10, 6), dpi=100) fig.tight_layout(pad=3.0) # 增加内边距，防止标签被截断 # 创建画布并添加到固定大小的框架 canvas = FigureCanvasTkAgg(fig, master=plot_frame) canvas.get_tk_widget().pack(side=tk.TOP, fill=tk.BOTH, expand=True) # 添加工具栏，包含缩放、保存等功能 toolbar_frame = ttk.Frame(plot_frame) toolbar_frame.pack(side=tk.BOTTOM, fill=tk.X) toolbar = NavigationToolbar2Tk(canvas, toolbar_frame) toolbar.update() # 启用紧凑布局，并设置自动调整以使图表完全显示 def on_resize(event): fig.tight_layout() canvas.draw_idle() # 添加图表交互功能 canvas.mpl_connect('resize_event', on_resize) canvas.mpl_connect('scroll_event', on_scroll) canvas.mpl_connect('motion_notify_event', update_cursor) # 添加鼠标拖动功能 def on_press(event): if event.inaxes != ax: return canvas.get_tk_widget().config(cursor="fleur") ax._pan_start = (event.x, event.y, event.xdata, event.ydata) def on_release(event): ax._pan_start = None canvas.get_tk_widget().config(cursor="") canvas.draw_idle() def on_motion(event): if not hasattr(ax, '_pan_start') or ax._pan_start is None: return if event.inaxes != ax: return start_x, start_y, x_data, y_data = ax._pan_start dx = event.x - start_x dy = event.y - start_y # 获取当前视图 xlim = ax.get_xlim() ylim = ax.get_ylim() # 计算图表坐标系中的移动 x_scale = (xlim[1] - xlim[0]) / canvas.get_tk_widget().winfo_width() y_scale = (ylim[1] - ylim[0]) / canvas.get_tk_widget().winfo_height() # 更新视图 ax.set_xlim(xlim[0] - dx * x_scale, xlim[1] - dx * x_scale) ax.set_ylim(ylim[0] + dy * y_scale, ylim[1] + dy * y_scale) # 更新拖动起点 ax._pan_start = (event.x, event.y, event.xdata, event.ydata) canvas.draw_idle() # 连接鼠标事件 canvas.mpl_connect('button_press_event', on_press) canvas.mpl_connect('button_release_event', on_release) canvas.mpl_connect('motion_notify_event', on_motion) # 更新历史数据显示函数 def display_history(): try: ax.clear() end_date = df['DateTime'].max() start_date = max(df['DateTime'].min(), end_date - timedelta(days=60)) hist_data = df[(df['DateTime'] >= start_date) & (df['DateTime'] <= end_date)] if len(hist_data) == 0: status_label.config(text="警告: 没有可用的历史数据") return # 创建双y轴图表 ax2 = None has_water_level = 'water_level' in hist_data.columns and 'water_level_smooth' in hist_data.columns if has_water_level: ax2 = ax.twinx() # 绘制数据 ax.plot(hist_data['DateTime'], hist_data['downstream_smooth'], label='一取水(下游)盐度', color='blue', linewidth=1.5) ax.plot(hist_data['DateTime'], hist_data['upstream_smooth'], label='青龙港(上游)盐度', color='purple', linewidth=1.5, alpha=0.7) # 设置边界，确保有一致的视图 y_min = min(hist_data['downstream_smooth'].min(), hist_data['upstream_smooth'].min()) * 0.9 y_max = max(hist_data['downstream_smooth'].max(), hist_data['upstream_smooth'].max()) * 1.1 ax.set_ylim(y_min, y_max) # 如果有水位数据，在第二个y轴上绘制 if ax2 is not None and has_water_level: try: # 检查水位数据是否有足够的非NaN值 valid_water_level = hist_data['water_level_smooth'].dropna() if len(valid_water_level) > 10: # 至少有10个有效值 ax2.plot(hist_data['DateTime'], hist_data['water_level_smooth'], label='长江水位', color='green', linewidth=1.5, linestyle='--') ax2.set_ylabel('水位 (m)', color='green') ax2.tick_params(axis='y', labelcolor='green') # 创建组合图例 lines1, labels1 = ax.get_legend_handles_labels() lines2, labels2 = ax2.get_legend_handles_labels() ax.legend(lines1 + lines2, labels1 + labels2, loc='best') else: print("水位数据有效值不足，跳过水位图") ax.legend(loc='best') except Exception as e: print(f"绘制水位数据时出错: {e}") ax.legend(loc='best') else: ax.legend(loc='best') # 设置标签和标题 ax.set_xlabel('日期') ax.set_ylabel('盐度') ax.set_title('历史数据对比') # 使用紧凑布局并绘制 fig.tight_layout() # 使用多种方法确保图像显示 plt.close(fig) # 关闭旧的 fig.canvas.draw() fig.canvas.flush_events() plt.draw() except Exception as e: status_label.config(text=f"显示历史数据时出错: {str(e)}") import traceback traceback.print_exc() display_history() root.mainloop() def resample_to_hourly(df): """ 将分钟级数据重采样为小时级数据，计算每小时的平均值 """ try: # 确保DateTime是索引 if 'DateTime' in df.columns: df = df.set_index('DateTime') # 获取所有数值列 numeric_columns = df.select_dtypes(include=[np.number]).columns.tolist() # 按小时重采样，计算平均值 hourly_df = df[numeric_columns].resample('H').mean() # 重置索引，将DateTime作为列 hourly_df = hourly_df.reset_index() print(f"数据已从分钟级重采样为小时级，原始数据行数: {len(df)}，重采样后行数: {len(hourly_df)}") return hourly_df except Exception as e: print(f"重采样数据异常: {e}") return df # ------------------------------- # 主程序入口：加载数据、添加特征、生成延迟特征后启动GUI # ------------------------------- def save_processed_data(df, filename='processed_data.pkl'): try: df.to_pickle(filename) print(f"已保存处理后的数据到 {filename}") return True except Exception as e: print(f"保存数据失败: {e}") return False def load_processed_data(filename='processed_data.pkl'): try: if os.path.exists(filename): df = pd.read_pickle(filename) print(f"已从 {filename} 加载处理后的数据") return df else: print(f"找不到处理后的数据文件 {filename}") return None except Exception as e: print(f"加载数据失败: {e}") return None # 删除旧的处理数据（如果存在），以应用修复后的代码 if os.path.exists('processed_data.pkl'): try: os.remove('processed_data.pkl') print("已删除旧的处理数据缓存，将使用修复后的代码重新处理数据") except Exception as e: print(f"删除缓存文件失败: {e}") # 删除旧的模型文件（如果存在） if os.path.exists('salinity_model.pkl'): try: os.remove('salinity_model.pkl') print("已删除旧的模型文件，将重新训练模型") except Exception as e: print(f"删除模型文件失败: {e}") # 尝试加载处理后的数据，如果不存在则重新处理 processed_data = load_processed_data() if processed_data is not None: df = processed_data else: # 添加长江液位数据作为参数 df = load_data('青龙港1.csv', '一取水.csv', '长江液位.csv', '大通流量.csv') if df is not None: # 添加时间特征 df['hour'] = df['DateTime'].dt.hour df['weekday'] = df['DateTime'].dt.dayofweek df['month'] = df['DateTime'].dt.month # 添加农历特征 df = add_lunar_features(df) # 添加延迟特征上游到下游3-5天，暂时每12小时为一个节点，根据效果后续再调整 # delay_hours = [1,2,3,4,6,12,24,36,48,60,72,84,96,108,120] delay_hours = [72,84,96,108,120] df = batch_create_delay_features(df, delay_hours) # 添加统计特征 df['mean_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_up'] = df['upstream_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).std() df['max_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).max() df['min_1d_up'] = df['upstream_smooth'].rolling(window=24, min_periods=1).min() df['mean_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_down'] = df['downstream_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).std() df['max_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).max() df['min_1d_down'] = df['downstream_smooth'].rolling(window=24, min_periods=1).min() # 添加水位统计特征（如果水位数据存在） if 'water_level' in df.columns: # 首先创建水位平滑特征 if 'water_level_smooth' not in df.columns: df['water_level_smooth'] = df['water_level'].rolling(window=24, min_periods=1, center=True).mean() df['water_level_smooth'] = df['water_level_smooth'].fillna(df['water_level']) # 添加水位统计特征 df['mean_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).mean() df['mean_3d_water_level'] = df['water_level_smooth'].rolling(window=72, min_periods=1).mean() df['std_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).std() df['max_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).max() df['min_1d_water_level'] = df['water_level_smooth'].rolling(window=24, min_periods=1).min() # 计算水位变化率 df['water_level_change_1h'] = df['water_level_smooth'].diff() df['water_level_change_24h'] = df['water_level_smooth'].diff(24) # 计算水位与盐度的相关特征 df['water_level_sal_ratio'] = df['water_level_smooth'] / df['downstream_smooth'] print("水位特征已添加") # 添加其他特征 df = generate_features(df) # 将数据重采样为小时级 df = resample_to_hourly(df) # 保存处理后的数据 df.to_csv('merged_data_hour.csv', index=False) print(f"Merged data saved to 'merged_data_hour.csv' successfully") save_processed_data(df) if df is not None: run_gui() else: print("数据加载失败，无法运行预测。")