In this study a neural network for state of charge (SoC) estimation of a lead acid battery was developed. The network structure is based on a nonlinear autoregressive neural network (NARX). This network structure assures consistent SoC behaviour over time without oscillations. The network has three hidden layers and a feedback loop for the SoC. The inputs of the network are battery current, -voltage, derivative of the battery voltage and previous predicted SoC values. The output is the current SoC of the battery. To train the network a real-life dataset from a stationary battery storage system was used. The dataset was analysed, and a database of charge and discharge segments was created. The training dataset was developed by randomly combining the charge and discharge segments. The resulting current profiles are used to cycle a lead-acid cell under laboratory conditions to obtain the resulting voltage and SoC-responses for the training dataset. To obtain a reference SoC, the profile is designed to end the charge and discharge at 100% and 0% SoC respectively. In this case, Ah-counting can be used as a reference. Several charge/discharge profiles were created with the method described above and cross-validation was carried out to train and validate the neural network performance. The overall error of the dataset is less than two percent SoC in laboratory measurements. For further validation, two partial state of charge (PSoC) profiles were created using the database. During the profiles, the battery is cycled around 60% SoC for two days before it is discharged to 0% SoC. For these measurements, no reference SoC can be calculated for validation because the charge efficiency of lead acid batteries is hard to specify and changes with SoC due to non-linearity of the side reactions. Therefore, the measurements are validated at the end of discharge at 1.8 V, which corresponds to 0% SoC. The SoC of the NARX network at the reference point for the two profiles is 1.6% and 5.3% respectively. In comparison, the SoC was also determined for the two profiles using Ah-counting with current dependent capacity. This method yields 7.5% and 9.7% SoC respectively. In conclusion, the NARX network is up to 6% more accurate than Ah-counting but still yields maximum errors up to 5% SoC. In the future, a method will be developed to test and validate the algorithm using actual field data.