Objective: Electrical equipment is highly vulnerable to seismic hazards. Acquiring accurate seismic response data is crucial for post-earthquake damage assessment and emergency decision-making. Although conventional contact sensors are effective for capturing such data, they cannot be widely deployed on the equipment body due to monitoring constraints. Therefore, seismic response prediction methods based on time-series neural networks must be developed. Most existing studies have emphasized the optimization of neural network architectures, and dataset construction strategies have not been systematically and sufficiently investigated. Dataset construction directly influences the fitting accuracy and generalization ability of predictive models, ultimately determining the overall predictive performance. In this study, the effects of three key elements of dataset construction-ground-motion selection, amplitude scaling of records, and sample-size configuration-on the performance of time-series models were evaluated, and a scientifically grounded dataset construction workflow was proposed. Methods: A 500-kV transformer bushing was selected as the case study. A refined finite element model was developed and validated by shaking-table tests to determine its accuracy in terms of dynamic characteristics and response behavior, and the acceleration at the top oil reservoir was chosen as the prediction target. Two ground-motion selection strategies were adopted for dataset construction: spectrum-matched records and random selection constrained only by site type. Four amplitude scaling strategies were examined: conventional random, conventional fixed, extended-range random, and extended-range fixed scaling. Five sample-size levels of 80, 100, 120, 140, and 160 records were also configured to form multiple strategy combinations. A recursive long short-term memory neural network was used as the representative prediction model. Its performance was assessed based on mean squared error and peak response error, and repeated sampling and multiple independent training runs were performed to mitigate stochastic variability. Results: Spectrum-matched selection outperformed random selection based solely on the site type, yielding lower overall prediction errors in seismic response time series. Fixed scaling was superior to random scaling, and the introduction of extended-range scaling further enhanced the peak prediction accuracy. Although commonly used, random scaling considerably reduced the overall and peak prediction performance of the model and was not recommended for seismic response time-series prediction. Increasing the number of training samples improved the model accuracy, but marginal gains were observed at a sample size of 120-140 records. The combination of spectrum-matched selection and extended-range fixed scaling was the most effective strategy. Comparative tests with representative ground-motion records further confirmed that this strategy surpassed commonly used empirical approaches in terms of fitting accuracy, peak prediction capability, and training stability; it also enabled more accurate capture of abrupt response transitions and reduced phase errors. Conclusions: A recommended dataset construction workflow is proposed for the time-series prediction of electrical equipment modeled as linear elastic systems. The proposed process integrates finite element modeling and validation, spectrum-matched ground-motion selection, extended-range fixed scaling, and balanced sample-size configuration. The findings confirm that this workflow considerably improves both the prediction accuracy and overall stability of the model, offering systematic methodological support and practical engineering guidance for post-earthquake emergency assessment and response monitoring of electrical equipment. This approach can also be extended to other structural systems where dataset construction critically affects the model performance.