Deep Learning for Power System Load Modeling: A BiLSTM Approach Validated on Real Transmission System Events

Accurate modelling of electrical load behaviour remains a long-standing challenge in power system security assessment, since the load is the only system component whose parameters cannot be declared by manufacturers and must instead be inferred from observed system response. Inaccurate load representation is a common root cause of voltage-instability misdiagnosis, and the problem is aggravated by the ongoing integration of distributed energy resources and power-electronicinterfaced consumption. This paper presents a deep-learning approach, based on a bidirectional long short-term memory (BiLSTM) network, for identifying the parameters of a composite ZIP (constant impedance Z, constant current I, constant power P) plus induction-motor load model from short time series of bus voltages and currents recorded during voltage disturbances. The network is trained on a synthetic dataset of 30,000 three-phase fault simulations, covering a broad range of static–dynamic load compositions, ZIP shares, and rotor parameters. The trained model is first evaluated on an unseen synthetic test set and then fine-tuned on three real fault recordings captured in a Croatian 110/35 kV substation and tested on two further unseen real fault recordings at the same substation. The BiLSTM recovers the dynamic load parameters with R2 > 0.97 on the synthetic test set, identifies the constant-power and constant-impedance ZIP shares with good accuracy, and yields physically plausible parameter estimates on the real recordings, in line with the assumed composite-load characteristics of the supplied area. Themethod is non-invasive, relies only on routinely available fault records, and directly supports advanced N−1 contingency analysis performed by transmission system operators.