LSTM-CNN多输出预测加optuna自动调参神经网络神经元个数dropout和batch-size

import os import matplotlib.pyplot as plt import numpy as np import optuna import pandas as pd from numpy.random import seed from sklearn.metrics import mean_absolute_error from tensorflow.keras.layers import Convolution1D, Activation, Conv1D from tensorflow.keras.layers import Dense from tensorflow.keras.models import Sequential from tensorflow.keras.optimizers import Adam from tensorflow.python.framework.random_seed import set_random_seed from tensorflow.python.keras.layers import CuDNNLSTM, Flatten set_random_seed(11) seed(7) os.chdir(r'C:\Projects\LStM时序预测加自动调参') train_path = 'Data.csv' data = pd.read_csv(train_path) train = data.iloc[51:, :] test = data.iloc[:50, :] train.dropna(inplace=True) train = train.values test = test.values # 建立CNN+LSTM模型 def create_model(look_back, units1, conv1, units2, lr, weight_decay): model = Sequential() model.add(CuDNNLSTM(units=units1, return_sequences=True, input_shape=(look_back, 3))) model.add(Convolution1D(conv1, 4, padding='same', strides=2)) model.add(Activation('relu')) model.add(Conv1D(units2, 4, strides=2, activation='relu', padding="same")) model.add(Flatten()) model.add(Dense(pred)) optimizer = Adam(learning_rate=lr, decay=weight_decay) model.compile(loss='mse', optimizer=optimizer, metrics=['mae']) return model def fit_size(train, y_train, test, test_y): from sklearn import preprocessing x_MinMax = preprocessing.MinMaxScaler() y_MinMax = preprocessing.MinMaxScaler() x = x_MinMax.fit_transform(train) y = y_MinMax.fit_transform(y_train) test_x = x_MinMax.transform(test) test_y = y_MinMax.transform(test_y) return x, y, test_x, test_y, y_MinMax # 预测第6小时,第12小时的把pred改成12 def create_dataset(dataset, train_y, TIME_STEPS, pred): dataX, dataY = [], [] for i in range(len(dataset) - TIME_STEPS - pred - 1): a = dataset[i:(i + TIME_STEPS), :] dataX.append(a) y = train_y[(i + TIME_STEPS + 1):(i + TIME_STEPS + pred + 1), 0].tolist() dataY.append(y) TrainX = np.array(dataX) Train_Y = pd.DataFrame(dataY).values return TrainX, Train_Y # 预测第6小时,第12小时的把pred改成12 def create_dataset1(dataset, train_y, TIME_STEPS, pred): dataX, dataY = [], [] for i in range(0, int(len(dataset) / 49)): x = dataset[i * 49:(i + 1) * 49, :] y49 = train_y[i * 49:(i + 1) * 49, :] for i in range(len(x) - TIME_STEPS - pred): a = x[i:(i + TIME_STEPS), :] dataX.append(a) y = y49[(i + TIME_STEPS + 1):(i + TIME_STEPS + pred + 1), 0].tolist() dataY.append(y) TrainX = np.array(dataX) Train_Y = pd.DataFrame(dataY).values return TrainX, Train_Y # 调参方法 def objective(trial): weight_decay = trial.suggest_uniform('weight_decay', 0.009, 0.025) lr = trial.suggest_uniform('lr', 0.002, 0.04) conv1 = trial.suggest_int('conv1', 32, 50) units1 = trial.suggest_int('units1', 1, 16) units2 = trial.suggest_int('units2', 32, 64) drop1 = trial.suggest_uniform('drop1', 0.2, 0.4) drop2 = trial.suggest_uniform('drop2', 0.3, 0.5) epochs = trial.suggest_int('epochs', 10, 30) print([weight_decay, lr, conv1, units1, units2, drop1, drop2, epochs]) model = create_model(look_back, units1, conv1, units2, lr, weight_decay) model.fit(X_train, y_train, batch_size=batch_size, validation_split=0.1, epochs=epochs, shuffle=False, verbose=0) pred_y = model.predict(X_test, batch_size=batch_size) mae = mean_absolute_error(test_y, pred_y) print(mae) return mae # 使用6行之前的数据 look_back = 6 # 预测未来4行数据 pred = 4 batch_size = 64 # 输入的参量可以选择，不是全部输入,可以这样对比结果 y_train = train[:, 0].reshape(-1, 1) test_y = test[:, 0].reshape(-1, 1) X_train, y_train, X_test, test_y, train_y_MinMax = fit_size(train, y_train, test, test_y) X_train, y_train = create_dataset1(X_train, y_train, look_back, pred) X_test, test_y = create_dataset(X_test, test_y, look_back, pred) """ Run optimize. Set n_trials and/or timeout (in sec) for optimization by Optuna """ study = optuna.create_study(direction='minimize') study.optimize(objective, n_trials=50) print('Best trial number: ', study.best_trial.number) print('Best value:', study.best_trial.value) print('Best parameters: \n', study.best_trial.params) parameters = study.best_trial.params # parameters = {'weight_decay': 0.01290914, 'lr': 0.0015873, 'conv1': 48, 'units1': 15, 'units2': 44, # 'drop1': 0.3604710283, 'drop2': 0.35916614, 'epochs': 43} # {'weight_decay': 0.011544699692321633, 'lr': 0.03714850010671891, 'conv1': 48, # 'units1': 4, 'units2': 42, 'drop1': 0.20453530440333112, 'drop2': 0.42749144468700656, 'epochs': 24} # 找到的最好的参数 weight_decay = parameters['weight_decay'] lr = parameters['lr'] conv1 = parameters['conv1'] units1 = parameters['units1'] units2 = parameters['units2'] drop1 = parameters['drop1'] drop2 = parameters['drop2'] epochs = parameters['epochs'] model = create_model(look_back, units1, conv1, units2, lr, weight_decay) hist = model.fit(X_train, y_train, batch_size=batch_size, validation_split=0.1, epochs=epochs, shuffle=False, verbose=1) # 画训练集和验证集的准确率图 plt.plot(hist.history['mae'], label='mae') plt.plot(hist.history['val_mae'], label='val_mae') plt.legend() plt.show() # 画损失图 plt.plot(hist.history['loss'], label='loss') plt.plot(hist.history['val_loss'], label='val_loss') plt.legend() plt.show() # 预测测试集 pred_y = model.predict(X_test, batch_size=batch_size) print(pred_y) print('mae_score') mae = mean_absolute_error(test_y, pred_y) print(mae)