<p>Predicting short-term power prices accurately is crucial to preserving market stability and assisting market players in making well-informed decisions. Forecasting is made much more difficult by the extreme volatility and intricate relationships between a number of driving variables that affect power pricing in the Australian National Electricity Market (NEM). A hybrid forecasting framework that incorporates a Bidirectional Long Short-Term Memory (BiLSTM) network and an inverted Transformer (iTransformer) is developed in this research to address these issues. Using self-attention across the feature dimension and considering individual variables as independent tokens, the iTransformer reorganizes multivariate input data in the suggested framework. Cross-variable dependencies between past pricing, system load, and weather conditions may be effectively learned thanks to this architecture. A BiLSTM module then processes the feature-aware representations in order to capture both short-term fluctuations and longer-range temporal relationships by using bidirectional temporal information. A multi-window Isolation Forest approach is used to find and exclude anomalous observations before model training, and real operational data from the Australian NEM in 2019 is used for empirical assessment. To guarantee reliable performance, Bayesian search techniques are used to identify important model parameters. Experiments are conducted in comparison to a number of widely used benchmark models, such as the conventional iTransformer, BPNN, LSTM, BiLSTM, and ATT-LSTM. The findings demonstrate that the suggested approach consistently produces fewer prediction errors across a variety of assessment parameters. Specifically, the model shows enhanced capacity to capture abrupt changes in pricing. These results show that in very volatile power markets, the suggested framework offers a workable and trustworthy way to anticipate short-term electricity prices.</p>