With the diversification of electricity trading forms driven by distributed energy technologies, the continuous growth of blockchain’s chained data structure poses dual challenges to traditional B+ tree indexes in terms of query efficiency and storage costs. This paper proposes a sliding window-based learned index construction method (SW-LI). The method consists of two key components. First, block timestamp–height samples are selected using a sliding window and used to train a linear regression model that captures the timestamp-to-height mapping. Second, an adaptive window adjustment mechanism is introduced: when the prediction error within a window exceeds a threshold, the window is contracted to improve local fitting accuracy; otherwise, it is expanded to accelerate global index construction. Together, these components dynamically balance model accuracy and training efficiency. Experimental results demonstrate that when the block count increases from 5000 to 25,000, SW-LI improves index construction efficiency by 69.22%–88.22% compared to Anole. Under a 10,000-block scale, its prediction error is reduced by an average of 80% compared to Sliding Window Search-enhanced Online Gradient Descent (SWS-OGD), with a storage overhead of only 60 KB (25,000 blocks), validating the method’s ability to maintain query accuracy while significantly enhancing indexing efficiency. When the block contains 4000 transactions, the average total query latency of SW-LI is 46.15% lower than that of Anole, which is only 2.7% of the average query latency of SWS-OGD (i.e., approximately 37 times faster).
Qi et al. (Thu,) studied this question.