Marine safety and security, including military situational awareness, create a need for short-term ocean forecasts of the ocean state. At present, such forecasts are done using numerical models on high performance computers. These models, although robust, often provide information at relatively coarse resolution for local/regional application, are difficult to deploy rapidly, and require connection between the data centre and user in the field. However, the rapid development of data-driven methods has opened a possibility for statistical emulators that can be trained with numerical model data to produce similar information but can be evaluated on a laptop computer in the field in a matter of seconds. To this end, we demonstrate creating an emulator to predict thermocline depth, an important parameter for sonar weather, up to 10 days in advance in the Archipelago Sea, Baltic Sea. We find that for this purpose, multiple linear regression produces the best results as more complex architectures suffered from limited training data or limited memory in the training phase.