Seismic structural health monitoring (SHM) system typically covers a limited number of buildings in a region to measure their dynamic responses, while the responses of the other non-instrumented buildings remain unknown during seismic events. It is widely used to predict these responses by nonlinear time history analysis (NLTHA). In recent years, deep learning (DL) models have emerged as efficient surrogate models for NLTHA due to their remarkable accuracy and efficiency. However, models like long-short-term memory (LSTM) are limited in predicting seismic response sequences for individual buildings, hindering comprehensive regional assessments. To address this, we propose a novel DL model based on LSTM to reconstruct seismic response sequences across multiple buildings within the same cluster. The inputs of the proposed model consist of ground motion acceleration, seismic responses of instrumented indicator buildings, and building attributes of non-instrumented target buildings, with outputs providing the seismic response of these target buildings. The feasibility of the proposed model will be demonstrated in a hypothetical region with 31 high-rise reinforced concrete shear walls (1 indicator building and 30 target buildings). The building attributes include the number of stories and floor area. Each building will be simplified as a multi-degree of freedom (MDOF) shear model with a pinching hysteretic model to characterize the inter-story behavior. NLTHA will be conducted under a total of 65 ground motion records to obtain response sequences to generate training data. The results show that the proposed DL model achieves 93.5% accuracy when predicting roof acceleration response sequences of 30 non-instrumented buildings under 14 testing seismic waves. One prediction can be completed in 0.05 ms, capable of real-time application. Overall, the proposed DL model has the potential to enhance regional seismic risk assessment with a limited SHM system.