import sys import pandas as pd import numpy as np import matplotlib.pyplot as plt from sklearn.preprocessing import MinMaxScaler from sklearn.metrics import mean_absolute_error, mean_squared_error # Check numpy version if np.__version__ < '1.23.5': print(f"Warning: Current numpy version {np.__version__} may cause compatibility issues.") print("Please upgrade numpy to version 1.23.5 or higher.") try: import tensorflow as tf print(f"TensorFlow version: {tf.__version__}") except ImportError as e: print("Error importing TensorFlow:", e) sys.exit(1) # Check TensorFlow version if tf.__version__ < '2.10.0': print(f"Warning: Current TensorFlow version {tf.__version__} may cause compatibility issues.") print("Please upgrade TensorFlow to version 2.10.0 or higher.") from tensorflow.keras.models import Model from tensorflow.keras.layers import Input, Conv1D, LSTM, Dense, Layer from tensorflow.keras.callbacks import EarlyStopping # ================================ # 1. 数据预处理模块 # ================================ class DataPreprocessor: def __init__(self, file_path, target_col='Value'): # 使用正确的编码和列名 self.df = pd.read_csv(file_path, encoding='utf-8-sig', parse_dates=['DateTime'], index_col='DateTime') self.target_col = target_col self.scaler = MinMaxScaler(feature_range=(0, 1)) def preprocess(self, resample_freq='h'): # 使用小写的'h'代替大写的'H' """数据重采样与归一化""" # 只选择Value列进行处理 value_series = self.df[self.target_col] # 处理非等间隔采样 df_resampled = value_series.resample(resample_freq).mean() df_filled = df_resampled.fillna(method='ffill').fillna(method='bfill') # 双向填充 # 归一化处理 self.scaled_data = self.scaler.fit_transform(df_filled.values.reshape(-1, 1)) self.dates = df_filled.index return self.scaled_data, self.dates def create_sequences(self, data, look_back=72, pred_steps=120): """创建监督学习序列""" X, Y = [], [] for i in range(len(data) - look_back - pred_steps): X.append(data[i:(i + look_back)]) Y.append(data[(i + look_back):(i + look_back + pred_steps)]) return np.array(X), np.array(Y) # ================================ # 2. 模型构建模块 # ================================ class TemporalAttention(Layer): """时间注意力机制层""" def __init__(self, units): super(TemporalAttention, self).__init__() self.W1 = Dense(units) self.W2 = Dense(units) self.V = Dense(1) def call(self, encoder_output, lstm_output): lstm_output = tf.expand_dims(lstm_output, 1) score = self.V(tf.nn.tanh( self.W1(encoder_output) + self.W2(lstm_output))) attention_weights = tf.nn.softmax(score, axis=1) context_vector = attention_weights * encoder_output context_vector = tf.reduce_sum(context_vector, axis=1) return context_vector, attention_weights class SalinityPredictor: def __init__(self, look_back=72, pred_steps=120): self.look_back = look_back # 历史窗口（小时） self.pred_steps = pred_steps # 预测步长（5天=120小时） def build_model(self): """构建CNN-LSTM-Attention混合模型""" inputs = Input(shape=(self.look_back, 1)) # CNN特征提取 cnn = Conv1D(64, 3, activation='relu', padding='same')(inputs) cnn = Conv1D(32, 3, activation='relu', padding='same')(cnn) # LSTM时序建模 lstm_out = LSTM(128, return_sequences=True)(cnn) lstm_out = LSTM(64, return_sequences=False)(lstm_out) # 注意力机制 context_vector, _ = TemporalAttention(64)(cnn, lstm_out) # 输出层 outputs = Dense(self.pred_steps)(context_vector) self.model = Model(inputs=inputs, outputs=outputs) self.model.compile(optimizer='adam', loss='mse') return self.model def dynamic_split(self, data, dates, cutoff_date): """动态划分训练集""" cutoff_idx = np.where(dates <= cutoff_date)[0][-self.look_back] train_data = data[:cutoff_idx] return train_data def train(self, X_train, y_train, epochs=200, batch_size=32): """模型训练""" early_stop = EarlyStopping(monitor='val_loss', patience=20) history = self.model.fit( X_train, y_train, epochs=epochs, batch_size=batch_size, validation_split=0.2, callbacks=[early_stop], verbose=1 ) return history def predict(self, last_sequence): """递归多步预测""" predictions = [] current_seq = last_sequence.copy() for _ in range(self.pred_steps): pred = self.model.predict(current_seq[np.newaxis, :, :], verbose=0) predictions.append(pred[0][0]) # 只取第一个预测值 current_seq = np.roll(current_seq, -1, axis=0) current_seq[-1] = pred[0][0] # 使用单个预测值更新序列 return np.array(predictions) # ================================ # 3. 完整流程执行 # ================================ if __name__ == "__main__": # 参数配置 DATA_PATH = 'D:\opencv\.venv\一取水.csv' CUTOFF_DATE = '2024-12-20 00:00' # 用户指定分割时间点 LOOK_BACK = 72 # 3天历史数据 PRED_STEPS = 120 # 预测5天 # 数据预处理 preprocessor = DataPreprocessor(DATA_PATH) scaled_data, dates = preprocessor.preprocess() X, Y = preprocessor.create_sequences(scaled_data, LOOK_BACK, PRED_STEPS) # 模型构建 predictor = SalinityPredictor(LOOK_BACK, PRED_STEPS) model = predictor.build_model() model.summary() # 动态训练 train_data = predictor.dynamic_split(scaled_data, dates, pd.to_datetime(CUTOFF_DATE)) X_train, y_train = preprocessor.create_sequences(train_data, LOOK_BACK, PRED_STEPS) history = predictor.train(X_train, y_train) # 预测验证 cutoff_idx = np.where(dates <= pd.to_datetime(CUTOFF_DATE))[0][-1] last_seq = scaled_data[cutoff_idx-LOOK_BACK:cutoff_idx] # 使用分割点前的数据作为输入 scaled_pred = predictor.predict(last_seq) predictions = preprocessor.scaler.inverse_transform(scaled_pred.reshape(-1, 1)) # 结果可视化 true_dates = pd.date_range(start=pd.to_datetime(CUTOFF_DATE), periods=PRED_STEPS, freq='h') plt.figure(figsize=(15, 6)) # 绘制分割点前的历史数据 plt.plot(dates[cutoff_idx-PRED_STEPS:cutoff_idx], preprocessor.scaler.inverse_transform(scaled_data[cutoff_idx-PRED_STEPS:cutoff_idx]), 'b-', label='历史数据（分割点前）') # 绘制分割点后的实际数据 plt.plot(dates[cutoff_idx:cutoff_idx+PRED_STEPS], preprocessor.scaler.inverse_transform(scaled_data[cutoff_idx:cutoff_idx+PRED_STEPS]), 'g-', label='实际数据（分割点后）') # 绘制预测数据 plt.plot(true_dates, predictions, 'r--', label='预测数据') # 添加分割线 plt.axvline(x=pd.to_datetime(CUTOFF_DATE), color='k', linestyle='--', label='分割时间点') plt.title(f'盐度预测对比（{CUTOFF_DATE}后5天）') plt.xlabel('时间') plt.ylabel('盐度值') plt.legend() plt.grid(True) plt.xticks(rotation=45) plt.tight_layout() plt.show() # 性能指标 true_values = preprocessor.scaler.inverse_transform(scaled_data[cutoff_idx:cutoff_idx+PRED_STEPS]) mae = mean_absolute_error(true_values, predictions) rmse = np.sqrt(mean_squared_error(true_values, predictions)) print(f'验证指标 => MAE: {mae:.3f}, RMSE: {rmse:.3f}') # ================================ # 4. 模型保存与加载（可选） # ================================ # model.save('salinity_predictor.h5') # loaded_model = tf.keras.models.load_model('salinity_predictor.h5', custom_objects={'TemporalAttention': TemporalAttention})