💥1 概述
在本文中，我们提出了我们的策略和算法来管理微电网中的能量，更具体地说，在考虑消费者负荷、太阳能发电和动态电价的情况下，每小时与主电网的能量交易。
该算法的主要目标是在用户负荷、太阳能发电和电价波动的情况下，优化储能系统( ESS )的运行，最大化微电网的货币效益。除了货币收益外，该算法还考虑了需要保留的最小能量，因为这对于确保微电网中关键任务操作的连续性至关重要。研究分析了两种能量管理算法的性能：1 )具有预测未来知识的模型预测控制线性规划( MPCLPF )；2 )不具有未来知识的强化学习。在MPCLPF中，分析了不同的预测算法，并对最优预测算法进行了整合。下面是文章目录：
详细文章见第4部分。
📚2 运行结果
fig, ax = plt.subplots(1, 1, figsize = (7.5,5)) ax2 = ax.twinx() PV_plot = ax.step(np.arange(24), df.iloc[0:24,0], 'ro-', label = "PV") load_plot = ax.step(np.arange(24), df.iloc[0:24,1], 'b*-', label = "Load") price_plot = ax2.step(np.arange(24), df.iloc[0:24,2], 'k.-', label = "RTP") # Display all label in one box plots = PV_plot + load_plot + price_plot labels = [plot.get_label() for plot in plots] ax.legend(plots, labels, loc = 0) ax.set_xlabel("Hour") ax.set_ylabel("Power (kW)") ax2.set_ylabel("Price ($/ kWh)") plt.show()
fig, ax = plt.subplots(3, 2, figsize = (15, 15)) ax[0, 0].step(np.arange(len(x[:,0])), x[:,0]) ax[0, 0].set_xlabel("Hour") ax[0, 0].set_ylabel("PV (kW)") ax[0, 1].step(np.arange(len(x[0:24,0])), x[0:24,0]) ax[0, 1].set_xlabel("Hour") ax[0, 1].set_ylabel("PV (pu)") ax[1, 0].step(np.arange(len(x[:,1])), x[:,1]) ax[1, 0].set_xlabel("Hour") ax[1, 0].set_ylabel("Load (kW)") ax[1, 1].step(np.arange(len(x[0:24,1])), x[0:24,1]) ax[1, 1].set_xlabel("Hour") ax[1, 1].set_ylabel("Load (pu)") ax[2, 0].step(np.arange(len(x[:,2])), x[:,2]) ax[2, 0].set_xlabel("Hour") ax[2, 0].set_ylabel("Price ($/kWh)") ax[2, 1].step(np.arange(len(x[0:24,2])), x[0:24,2]) ax[2, 1].set_xlabel("Hour") ax[2, 1].set_ylabel("Price (pu)") plt.show()
plt.plot(lstm.history.history["loss"], "-*", label="training") plt.plot(lstm.history.history["val_loss"], "-o", label="validation") plt.xticks(np.arange(0, 20, 2), np.arange(0, 20, 2)) plt.xlabel("Epoch") plt.ylabel("MAE") plt.legend() plt.show()
encoder.load_weights(encoder.weights_dir) decoder.load_weights(decoder.weights_dir) y_train_pred, attentions = predict(x_train, y_train) print ("Training MAE: {:.4f} pu\n".format(mae(y_train[:, :, 0], y_train_pred[:, :, 0]))) fig = plt.figure(figsize=(24, 5)) for idx, i in enumerate([0, 1000, 2000, 3000]): ax = fig.add_subplot(1, 4, idx+1) ax.plot(y_train_pred[i], "-*", label="prediction") ax.plot(y_train[i, :, 0], "-o", label="actual") ax.set_xlabel("Hour") ax.set_ylabel("Power (pu)") ax.legend(loc=2) plt.show()
plt.plot(lstm.history.history["loss"], "-*", label="training") plt.plot(lstm.history.history["val_loss"], "-o", label="validation") plt.xticks(np.arange(0, 20, 2), np.arange(0, 20, 2)) plt.xlabel("Epoch") plt.ylabel("MAE") plt.legend() plt.show()
idx = -10 num_steps_display = timesteps_in attention = attention_weights attention = tf.squeeze(attention["decoder_layer1_block2"][idx:idx+1], axis=0) for head in range(0, num_heads): fig = plt.figure(figsize=(32,8)) spec = gridspec.GridSpec(ncols=90, nrows=100) top_ax = fig.add_subplot(spec[0:15, 15:75]) left_ax = fig.add_subplot(spec[25:, 0:10]) right_ax = fig.add_subplot(spec[25:, 15:]) top_ax.plot(x_train[idx, :num_steps_display, 0]) top_ax.set_xlim([0, num_steps_display]) top_ax.set_xticks(range(0, num_steps_display, 4)) top_ax.set_xticklabels(range(0, num_steps_display, 4)) left_ax.plot(decoder_input[idx, :, 0], range(0, timesteps_out)) left_ax.set_yticks(range(0, timesteps_out, 4)) left_ax.set_yticklabels(range(0, timesteps_out, 4)) left_ax.invert_yaxis() sns.heatmap(attention[head][:, :num_steps_display], cmap="viridis", ax=right_ax) right_ax.set_xticks(range(0, num_steps_display, 4)) right_ax.set_xticklabels(range(0, num_steps_display, 4)) right_ax.set_yticks(range(0, timesteps_out, 4)) right_ax.set_yticklabels(range(0, timesteps_out, 4)) plt.title("Head {}".format(head+1)) plt.show()
def get_resultplot(SOC_list, action_list, x, start_idx, end_idx): hours = end_idx - start_idx if hours == 24: plt.figure(figsize = (8,7)) plt.xticks(range(0, 24), range(1, 25)) else: plt.figure(figsize = (25,5)) plt.xticks(range(0, end_idx-start_idx, 24), range(1, end_idx-start_idx+1, 24)) plt.step(range(0, hours), SOC_list[start_idx:end_idx], "ro-", label = "SOC") plt.step(range(0, hours), x[start_idx:end_idx, 2], "bs-", label = "price") plt.step(range(0, hours), x[start_idx:end_idx, 0], "g*-", label = "pv") plt.step(range(0, hours), x[start_idx:end_idx, 1], "m--", label = "load") plt.bar(range(0, hours), action_list[start_idx:end_idx], facecolor = "w", edgecolor = "k", label = "action") plt.ylabel("SOC/ Normalized Price") plt.xlabel("Hour") plt.legend(loc=2) plt.show()
# Case 1 - Charged with PV not with grid to contain excess PV even the price is higher than average # Use the spare capacity to store PV # Not below the target SOC start_idx = len(SOC_list) - 192 end_idx = len(SOC_list) - 168 get_resultplot(SOC_list, action_list, x, start_idx, end_idx)
# Zoom of case 3 fig, ax = plt.subplots(1, 1, figsize = (8,6)) #ax2 = ax.twinx() ln1 = ax.step(range(0, 24), SOC_list[13079:13103], "ro-", label = "SOC") ln2 = ax.bar(range(0, 24), action_list[13079:13103], facecolor = "w", edgecolor = "k", label = "action") ln3 = ax.axhline(y = 0.5, linestyle = "--", label = "target SOC") ax.set_xlabel("Hour") ax.set_ylabel("SOC") lns = ln1 + [ln2] + [ln3] labs = [l.get_label() for l in lns] ax.legend(lns, labs, loc = 3) plt.xticks(range(0, 24), range(1, 25)) plt.show()
其余详细部分见第4部分。
🎉3 参考文献
部分理论来源于网络，如有侵权请联系删除。
