With the continuous advancement of communication systems towards high-speed and ultra-wideband, there has been a growing research focus on designing high-performance Raman amplifiers tailored to this progress. However, designing high-performance Raman amplifiers is challenging due to the complex non-linear relationship between output Raman gain, noise, and pump parameters. Traditional numerical optimization methods are not efficient in solving this problem. To address this, this paper proposes a second-order Raman fiber amplifier (RFA) gain and noise prediction model using convolutional neural network (CNN) and long-short term memory (LSTM) . The impact of different prediction model performances on the design of Raman fiber amplifiers is investigated, and to optimize the model using the sea horse optimizer (SHO) algorithm to accurately reflect the mapping relationship between pump parameters, fiber length, and the target gain and noise distribution. Experimental results show that the proposed model has a root mean square error of only 0.043 1 and 0.022 4 dB in gain and noise prediction, with an error between the predicted and target values below 0.25 dB and an average consumption time of less than 0.133 7 s. This design provides methods and ideas for the rapid design of Raman fiber amplifiers in the future.