models/custom_models.py

import numpy as np
import pandas as pd
import tensorflow as tf
import os
import tempfile
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import r2_score, root_mean_squared_error
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Bidirectional, GRU, Dense
from tensorflow.keras.callbacks import EarlyStopping
from statsmodels.tsa.arima.model import ARIMA
from statsmodels.tsa.holtwinters import ExponentialSmoothing
from prophet import Prophet

def prepare_data(df, lag=10):
    values = df.values.astype("float32")
    timestamps = df.index
    X, y, time = [], [], []

    for i in range(len(values) - lag):
        X.append(values[i : i + lag])
        y.append(values[i + lag, 0])
        time.append(timestamps[i + lag])

    return np.array(X), np.array(y), np.array(time)

def scale_data(X, y):
    n_features = X.shape[-1]
    scaler_X = MinMaxScaler()
    X_scaled = scaler_X.fit_transform(X.reshape(-1, n_features)).reshape(X.shape)
    scaler_y = MinMaxScaler()
    y_scaled = scaler_y.fit_transform(y.reshape(-1, 1)).flatten()
    return X_scaled, y_scaled, scaler_X, scaler_y

def invert_prediction(scaler, pred):
    return scaler.inverse_transform(pred.reshape(-1, 1)).flatten()

def create_model(input_shape, model_type, neurons):
    model = Sequential()
    if model_type == "LSTM":
        model.add(LSTM(neurons, input_shape=input_shape))
    elif model_type == "BiLSTM":
        model.add(Bidirectional(LSTM(neurons), input_shape=input_shape))
    elif model_type == "GRU":
        model.add(GRU(neurons, input_shape=input_shape))
    else:
        raise ValueError(f"Unsupported model type: {model_type}")
    model.add(Dense(1))
    model.compile(loss="mse", optimizer="adam")
    return model

def run_forecast(
    df,
    model_type,
    is_multivariate,
    horizon,
    lag,
    neurons,
    epochs,
    batch_size,
    future_horizon=0,
    device="CPU",
    arima_order=(5,1,0)
):
    if device == "GPU":
        physical_devices = tf.config.list_physical_devices('GPU')
        if physical_devices:
            tf.config.experimental.set_memory_growth(physical_devices[0], True)

    if model_type in ["ARIMA", "ExponentialSmoothing", "Prophet"]:
        return run_classical_forecast(df, model_type, horizon, future_horizon, arima_order)

    X, y, time = prepare_data(df, lag=lag)
    train_size = len(X) - horizon
    X_train, X_test = X[:train_size], X[train_size:]
    y_train, y_test = y[:train_size], y[train_size:]
    X_train, y_train, scaler_X, scaler_y = scale_data(X_train, y_train)
    X_test, y_test, _, _ = scale_data(X_test, y_test)

    model = create_model((lag, X.shape[-1]), model_type, neurons)

    early_stop = EarlyStopping(monitor="loss", patience=10, restore_best_weights=True)
    model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, verbose=1, shuffle=False, callbacks=[early_stop])

    y_test_pred = invert_prediction(scaler_y, model.predict(X_test))
    y_test = invert_prediction(scaler_y, y_test)

    future_pred = []
    if future_horizon > 0:
        last_input = X_test[-1]
        for _ in range(future_horizon):
            pred_scaled = model.predict(last_input.reshape(1, lag, -1))
            pred_inv = invert_prediction(scaler_y, pred_scaled)
            future_pred.append(pred_inv[0])
            next_input = np.roll(last_input, -1, axis=0)
            next_input[-1, 0] = pred_scaled
            last_input = next_input

    rmse = root_mean_squared_error(y_test, y_test_pred)
    r2 = r2_score(y_test, y_test_pred)
    metrics = f"Test RMSE: {rmse:.3f}, Test R2: {r2:.3f}"

    export_df = pd.DataFrame({"Test_Actual": y_test, "Test_Predicted": y_test_pred})
    export_path = os.path.join(tempfile.gettempdir(), "forecast_result.csv")
    export_df.to_csv(export_path, index=False)

    future_path = ""
    if future_pred:
        future_path = os.path.join(tempfile.gettempdir(), "forecast_future.csv")
        pd.DataFrame({"Future_Forecast": future_pred}).to_csv(future_path, index=False)

    fig = plt.figure(figsize=(12, 6))
    plt.plot(y_test, label="Test Actual")
    plt.plot(y_test_pred, label="Test Predicted", linestyle="--")
    if future_pred:
        plt.plot(range(len(y_test), len(y_test) + future_horizon), future_pred, label="Future Forecast", linestyle="-.")
    plt.title("Forecast Result")
    plt.xlabel("Time Step")
    plt.ylabel("Value")
    plt.legend()

    return y_test_pred, y_test, future_pred, fig, metrics, export_path, future_path

def run_classical_forecast(df, model_type, horizon, future_horizon=0, arima_order=(5,1,0)):
    ts = df.iloc[:, 0]
    ts.index = pd.to_datetime(df.index)

    if model_type == "ARIMA":
        model = ARIMA(ts, order=arima_order)
        model_fit = model.fit()
        forecast = model_fit.forecast(steps=horizon)
    elif model_type == "ExponentialSmoothing":
        model = ExponentialSmoothing(ts, trend="add", seasonal=None)
        model_fit = model.fit()
        forecast = model_fit.forecast(horizon)
    elif model_type == "Prophet":
        df_prophet = ts.reset_index()
        df_prophet.columns = ["ds", "y"]
        model = Prophet()
        model.fit(df_prophet)
        future = model.make_future_dataframe(periods=horizon)
        forecast_df = model.predict(future)
        forecast = forecast_df["yhat"].iloc[-horizon:].values
    else:
        raise ValueError(f"Model type {model_type} not supported.")

    actual = ts[-horizon:].values
    future_forecast = None

    if future_horizon > 0:
        if model_type == "Prophet":
            future = model.make_future_dataframe(periods=horizon + future_horizon)
            forecast_df = model.predict(future)
            future_forecast = forecast_df["yhat"].iloc[-future_horizon:].values
        else:
            future_forecast = model_fit.forecast(horizon + future_horizon)[-future_horizon:]

    rmse_val = np.sqrt(root_mean_squared_error(actual, forecast))
    r2_val = r2_score(actual, forecast)
    metrics = f"Test RMSE: {rmse_val:.3f}, Test R2: {r2_val:.3f}"

    export_df = pd.DataFrame({"Test_Actual": actual, "Test_Predicted": forecast})
    export_path = os.path.join(tempfile.gettempdir(), f"{model_type.lower()}_forecast_result.csv")
    export_df.to_csv(export_path, index=False)

    future_path = ""
    if future_forecast is not None:
        future_path = os.path.join(tempfile.gettempdir(), f"{model_type.lower()}_forecast_future.csv")
        pd.DataFrame({"Future_Forecast": future_forecast}).to_csv(future_path, index=False)

    fig = plt.figure(figsize=(12, 6))
    # Clean forecast for plotting
    if isinstance(forecast, pd.Series):
        forecast = forecast.values
    if isinstance(actual, pd.Series):
        actual = actual.values
    plt.plot(actual, label="Test Actual")
    plt.plot(forecast, label="Test Predicted", linestyle="--")
    if future_forecast is not None:
        plt.plot(range(len(actual), len(actual) + len(future_forecast)), future_forecast, label="Future Forecast", linestyle="-.")
    plt.title(f"{model_type} Forecast Result")
    plt.legend()

    return forecast, actual, future_forecast, fig, metrics, export_path, future_path