Abstract: Accurate forecasting of spot prices is crucial for market trading and grid dispatch. However, influenced by multiple intertwined factors, spot price series that exhibit high non-stationarity and complex periodic patterns, posing challenges for price forecasting. It is difficult for existing models to capture the intricate dynamics of price fluctuations, including the local and global characteristics of the price series and the complex patterns of multi-frequency signals. This paper proposes a heterogeneous deep learning ensemble prediction method that integrates the reconstruction-based secondary decomposition-ensemble(RSDE) framework and Kolmogorov- Arnold networks(KAN) algorithm, aiming to improve electricity price prediction accuracy by extracting local and global characteristics from different frequency domain signals comprehensively. First, the RSDE framework is introduced to make the original price signal into reconstructed sub-signals and related secondary decomposition sub-signals in high, medium, and low-frequency domains. Second, the KAN is incorporated into the deep learning model, applying prediction strategies from the RSDE framework to model and predict the reconstructed sub-sequences using various models. Finally, final prediction is obtained by integrating the prediction results of all reconstructed signals using adaptive weighted regression. An empirical analysis using electricity spot market price data from Australia demonstrates the priority of the proposed model, which outperforms the benchmark models in capturing complex fluctuation features.