In Li-ion battery applications, effective energy management relies heavily on accurate knowledge of the state of charge (SOC). As SOC cannot be directly measured, it must be estimated using several methods. Deep learning has emerged as one of the most widely used approaches in machine learning. However, in cases where the input data exhibit limited variation over time and consist of low-dimensional features, deep learning models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) may tend toward overfitting. To address this, deep learning algorithms such as long short-term memory (LSTM) have been focused on for SOC prediction. Nevertheless, the current-voltage behavior of Li-ion cells varies significantly under different operating conditions, such as charging, discharging, and idle states. This variability negatively impacts the performance of conventional LSTM models. To overcome this limitation, this study proposes a parallel LSTM architecture composed of three distinct models, each tailored to a specific battery operating condition. Both the proposed and conventional models were evaluated using various standardized driving cycles. Mean absolute error, mean squared error, and boxplot analysis were employed for performance comparison. Across all metrics, the proposed method consistently outperformed the standard model. The best mean absolute error result was achieved with the proposed method, at 0.75% under the LA92 driving cycle. These results demonstrate the effectiveness of the proposed approach in accurately and reliably estimating SOC in dynamic battery applications.