Predictive monitoring and anomaly detection in all the complex systems using all of the different types Multivariate Time Series (MTS) data is crucial. With the growing of the reliance on various platforms like cyber physical systems, early fault detection is crucial, but also challenging due to the high data dimensionality and lack of labeled anomalies in the respective datasets. The respective paper proposes an LSTM based forecasting model to learn normal system behavior and compute deviation based anomaly scores with out the need of respective labeled data. The forecasting component learns all of the temporal patterns and inter variable dependencies from historical MTS datasets which is used to predict future observations. The anomaly detection module then quantifies the deviation between the both the parameters which is the predicted and actual values to compute the respective anomaly score and also which serves as an indicator of potential faults or operational shifts in the respective model. Experiments on the UCI online retail and NAB dataset show strong results in effective anomaly detection (F1 score = 0.90 and accuracy = 97%), outperforming methods. The system provides interpretable early warnings and is scalable across domains, making it a practical solution for real-world monitoring and maintenance applications.