In-X subnetworks (SN) face challenges in achieving heterogeneous extreme requirements such as 0.1 ms latency and 6-to-8 nines reliability due to high and dynamic interference caused by mobility, channel characteristics, and limited resources in ultra-dense environments. We propose interference prediction techniques, M-to-M, 1-to-1, and federated learning (FL)-based long short-term memory (LSTM) that help in learning the complex and potentially non-linear relationship from a measured interference power vector. This results in significantly lower prediction error as compared to baseline methods. The FL-based predictor enhances robustness by leveraging knowledge from other sensor/actuator pairs' learning by achieving up to 4.5 dB lower prediction error as compared to M-to-M, and one-to-one LSTM predictors in the Sensor-actuator (SA) pair.