我试图通过'LSTM序列到序列'模型来预测时间序列数据 . 我用过keras . 我应该在模型中更改哪些内容以提高准确性？

  • 输入：过去4个时间步

  • 输出：未来2步

  • 特征数量：5

  • 数据维度：

train_x (1308, 4, 5)
train_y (1308, 2, 5)
test_x (118, 4, 5)
test_y (118, 2, 5)

结果： [mse : 0.021793483835408241, accuracy : 0.54661016696590492]

型号定义：

def fit_model(n_cells):
    model=Sequential()
    model.add(LSTM(n_cells, input_shape=(4,5)))
    model.add(Dense(n_cells))
    model.add(RepeatVector(2))  
    model.add(LSTM(n_cells,  input_shape=(10,5), return_sequences=True))
    model.add(TimeDistributed(Dense(5,  activation='linear')))  
    model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])
    print(model.summary())
    hist=History()
    for i in range(100):
    hist=model.fit(train_x, train_y, batch_size=32, epochs=1, validation_split=0.33, shuffle=False)
    model.reset_states()
    loss=model.evaluate(test_x, test_y, verbose=0)
    print(loss)

cells = [150]
for value in cells:
    fit_model(value)
keras lstm