df = pd.read_csv('/workspaces/Data-Miscellany-Forum/src/多变量时序预测系列-股票预测/多输入+单输出+单步/data.csv', parse_dates=["Date"], index_col=[0])df = pd.DataFrame(df)var_num = len(df.columns)print(df)

test_split=round(len(df)*0.20)df_for_training=df[:-test_split]df_for_testing=df[-test_split:]

scaler = MinMaxScaler(feature_range=(0, 1))df_for_training_scaled = scaler.fit_transform(df_for_training)df_for_testing_scaled=scaler.transform(df_for_testing)

train_dataset = TimeSeriesDataset(df_for_training_scaled, seq_len=seq_len, pred_len=pred_len)test_dataset = TimeSeriesDataset(df_for_testing_scaled, seq_len=seq_len, pred_len=pred_len)train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)

class LSTMRegressor(nn.Module):    def __init__(self, input_size=5, hidden_size=50, num_layers=1, output_size=1, dropout=0.2):        """        input_size: 输入特征维度（多变量个数）        hidden_size: LSTM隐藏层维度        num_layers: LSTM层数        """        super(LSTMRegressor, self).__init__()        self.hidden_size = hidden_size        # LSTM层        self.lstm = nn.LSTM(            input_size,             hidden_size,             num_layers,             batch_first=True)        # 全连接层        self.fc = nn.Sequential(            nn.Linear(hidden_size, hidden_size // 2),            nn.ReLU(),            nn.Dropout(dropout),            nn.Linear(hidden_size // 2, output_size)        )    def forward(self, x):        """        x: (batch_size, seq_length, input_dim)        return: (batch_size, output_dim)        """        lstm_out, (hidden, cell) = self.lstm(x)  # lstm_out: (batch, seq, hidden)        last_hidden = lstm_out[:, -1, :]  # (batch, hidden)取最后一个时间步的隐藏状态（或可用 hidden[-1]）        out = self.fc(last_hidden)  # (batch, output_dim)        return out

model = LSTMRegressor(input_size=var_num, hidden_size=hidden_size, num_layers=num_layers, output_size=pred_len, dropout=dropout).to(device)print("开始训练模型...")loss_history, train_preds, train_trues = train_model(model, train_loader, num_epochs=num_epochs, learning_rate=learning_rate, device=device)# 决策树基于LSTM残差训练residuals = np.array(train_trues).reshape(-1, 1)- np.array(train_preds).reshape(-1, 1)X_tree_train = train_preds.reshape(-1, 1)model_tree = DecisionTreeRegressor(max_depth=3)model_tree.fit(X_tree_train, residuals)

preds, trues = evaluate_model(model, test_loader, device=device)X_tree_test = preds.reshape(-1, 1)residual_preds = model_tree.predict(X_tree_test)final_preds = preds + residual_preds

plt.plot(loss_history, marker='o', color='dodgerblue', linestyle='-', linewidth=2)plt.title("Training Loss Curve")plt.xlabel("Epoch")plt.ylabel("MSE Loss")plt.tight_layout()plt.savefig('output_image1.png', dpi=300, format='png')plt.show()

plt.plot(trues, label="True Values", color='limegreen')plt.plot(preds, label="Predicted Values", color='crimson')plt.title("True vs. Predicted Values")plt.xlabel("Sample Index")plt.ylabel("Trend Value")plt.legend()plt.tight_layout()plt.savefig('output_image2.png', dpi=300, format='png')plt.show()

plt.plot(trues, label="True Values", color='limegreen')plt.plot(final_preds, label="Predicted Values", color='crimson')plt.title("True vs. Predicted Values")plt.xlabel("Sample Index")plt.ylabel("Trend Value")plt.legend()plt.tight_layout()plt.savefig('output_image3.png', dpi=300, format='png')plt.show()

sns.histplot(trues - preds, color='red', label='LSTM Residuals', kde=True)sns.histplot(trues - final_preds, color='green', label='Hybrid Residuals', kde=True)plt.title("Residual Distribution: LSTM vs Hybrid")plt.legend()plt.tight_layout()plt.savefig('output_image4.png', dpi=300, format='png')plt.show()

def evaluate_metrics(y_true, y_pred):    y_true = y_true.reshape(-1)    y_pred = y_pred.reshape(-1)    eps = 1e-8    mse = np.mean((y_true - y_pred)**2)    rmse = np.sqrt(mse + eps)    mae = np.mean(np.abs(y_true - y_pred))    mape = np.mean(np.abs((y_true - y_pred)/(np.abs(y_true)+eps)))    smape = np.mean(2*np.abs(y_true - y_pred)/(np.abs(y_true)+np.abs(y_pred)+eps))    # R2    ss_res = np.sum((y_true - y_pred)**2)    ss_tot = np.sum((y_true - np.mean(y_true))**2) + eps    r2 = 1 - ss_res/ss_tot    # Pearson    corr = np.corrcoef(y_true, y_pred)[0,1]    return dict(MSE=mse, RMSE=rmse, MAE=mae, MAPE=mape, sMAPE=smape, R2=r2, Corr=corr)