Boa noite.
Consegui resolver, era realmente o reshape pois um LSTM recebe na input [batch, timesteps, feature].
state = np.reshape(state, (state.shape[0], 1, state.shape[1])) e next_state = np.reshape(next_state, (next_state.shape[0], 1, next_state.shape[1]))
Obrigado!!
